The Legacy of State Farms on Ethiopia's Spatial Structural

Home , Agriculture in Ethiopia, Asaita, Economy of Ethiopia, Humera, Kafta Humera

W&M ScholarWorks

Undergraduate Honors Theses Theses, Dissertations, & Master Projects

5-2020

The Legacy of State Farms on Ethiopia’s Spatial Structural Transformation

Henry Larkin Young

Follow this and additional works at: https://scholarworks.wm.edu/honorstheses

Part of the Growth and Development Commons, and the Political Economy Commons

Recommended Citation Young, Henry Larkin, "The Legacy of State Farms on Ethiopia’s Spatial Structural Transformation" (2020). Undergraduate Honors Theses. Paper 1495. https://scholarworks.wm.edu/honorstheses/1495

This Honors Thesis is brought to you for free and open access by the Theses, Dissertations, & Master Projects at W&M ScholarWorks. It has been accepted for inclusion in Undergraduate Honors Theses by an authorized administrator of W&M ScholarWorks. For more information, please contact [email protected].

The Legacy of State Farms on Ethiopia’s Spatial Structural Transformation By Henry Larkin Young Thesis Advisors: Admasu Shiferaw and Philip Roessler Executive Committee: Berhanu Abegaz 5/6/2020

Abstract: This paper analyzes the determinants of Ethiopian commercial farm and agroprocessing agglomeration patterns. State farm investments during the Derg regime (1974- 1991) altered the geospatial distribution of commercial farmland, and concurrently agroprocessing production. Agglomeration patterns have stronger relationships with horticulture state farms due to infrastructure investments implemented by state farm planners. Agroprocessing firms also gain productivity advantages by sourcing inputs domestically, but recent value added growth in the agroprocessing sector was not attributed to commercial farm production growth. Finally, without the government acting as the main coordinating agency to develop commercial farm infrastructure within low-lying regions, high-value cash crop and agroprocessing production will be constrained by historic state farm determinants.

Acknowledgments I would like to first and foremost thank my parents for their unwavering support. I also want to thank my thesis advisers for guiding me and working with me on the research paper. The 1693 Program, Charles Center and the Economics Department at the College of William & Mary gave me the platform to conduct Honors Thesis research in Ethiopia. Without their financial assistance, this project would not have been possible. Furthermore, I want to thank the archivists at the National Archives of Ethiopia and Ministry of Agriculture for helping me find novel datasets and also the staff at the Center for Geospatial Analysis that taught me how to make sense of these materials and use them for my analysis. Finally, I want to thank everyone who read preliminary drafts; your edits were instrumental in turning this project into a research paper.

Table of Contents

I. Introduction…………………………………………………………….5-9

II. Historical Background……………………………………..………..10-26

III. Conceptual Framework and Methodology……………….…………27-47

IV. Data Sources……………………………………………...…………48-59

V. Data Analysis………………………………………………………..60-89

VI. Conclusion…………………………………………………………..90-91

VII. Sources……………………………………………………………....92-97

VIII. Appendix……...……………………………………………………98-104

Introduction Many African countries must see a vast reduction in their agricultural workforce and improve agricultural labor productivity to promote structural change (Collier and Dercon 2014).

Urbanization is an integral step in this process as smallholder farms migrate to cities to find employment in manufacturing (Henderson & Wang, 2005). Yet, African urbanization trends are substantially different than other continents since they are not driven by structural transformation

(Gollin et al. 2015). Without value added activities driving urbanization trends and inadequate public infrastructure investment, spatial bottlenecks often arise in primary cities like the capital.

Primary cities have weaker linkages to rural economies since industries and services are less effective at reaching rural areas (Dorosh et al. 2013). Recent literature has posited that second city development can improve spatial integration in regions that have stronger linkages to agricultural heartlands (Christiaensen et al. 2013, Vandercasteelen et al. 2018). Rural road investments also promote second-city development by reducing transportation costs for manufacturing firms (Shiferaw et al. 2015). As well, agroprocessing led industrialization can catalyze spatial integration since agroprocessing (mainly food, beverage, textile, and leather processing) firms serve regional markets as consumer consumption bundles change and provide employment opportunities for nearby rural areas (Dorosh et al. 2018).

Ethiopia is no exception; with 85% of the labor force still employed in agriculture and more than 30% of all manufacturing firms located in Addis Ababa, the country is characterized by stark spatial inequality (AGSS 2016 and CSA LLMIS 2015). To promote agroprocessing industrialization and urbanization in second cities, the government has constructed agroprocessing industrial parks in the agricultural heartland (UNIDO 2018). The government hopes that extension services for smallholder farmers will improve agricultural productivity so

5 they can provide the quantity of raw inputs needed to spur agroprocessing agglomeration. This development strategy follows a dominant trend of excluding commercial farming from rural development policy. Collier and Dercon (2014) have questioned the overreliance on smallholder agriculture and argue for more complementary forms of commercial agriculture production that could also boost peasant holding productivity. Private commercial farming could play an essential role in transforming the agricultural sector by consolidating landholdings, providing off-farm employment for smallholders, and facilitating peasant holder farmers’ attachment to commercial value chains in dense urban areas (Hall et al. 2017). The commercial farm sector could also provide a steady supply of raw material for agroprocessing firms since they already provide 25% of the wheat and maize supply for domestic agroprocessing firms (UNDP 2011).

Understanding the spatial determinants of commercial farmland location choices are important for promoting new commercial farm enclaves. Commercial farm location choices are highly selective and vary by country (Glover and Jones 2016 and Hall et al. 2015). But country specific research on the underlying structure of the commercial farm sector and its relation to agroprocessing sector is sparse. At the continent level, Roessler et al. 2019 finds that colonial era cash crop economies altered the geospatial distribution of economic activity since colonial infrastructure investments for extractive plantation agriculture promoted agglomeration.

Although Ethiopia does not share this colonial history, they share the same spatial characteristics since commercial farming is heavily concentrated in select woredas (counties). Between 2011-

2015, only 36% of woredas had any commercial farmland (CSA Commercial Farm Report 2011-

2015). Lavers et al. (2016) explains the specific pathways in which the government has attempted to promote foreign direct investment (FDI) within commercial agriculture, mainly through state farm privatization and tax holidays for commercial farm investment within low-

6 lying regions, but we do not know their consequences. State farms were formed during the Derg regime (1974-1991) when they nationalized all private land holdings. State farm investments were meant to produce the necessary marketed surplus to feed urban populations and support agroprocessing industrialization, but these behemoth investment projects only marginally outperformed smallholder farmers (World Bank 1987). Once the current EPRDF administration came to power in 1991, state farms were privatized and converted into commercial farms (EPA

2018). Recently, the government has identified unpopulated, low-lying regions as suitable locations for commercial farming and has provided tax holidays to investors (Lavers et al. 2012).

Over 80% of allotted private commercial farmland has been in low-lying regions, but there have not been any follow up studies analyzing whether these investments promoted new commercial farm enclaves and agroprocessing agglomeration or whether historic determinants drive still drive agglomeration patterns.

I fill this gap in the literature by analyzing how historic determinants and present policy affects the commercial farm and agroprocessing sector. I analyze whether the present location choices of commercial farms and agroprocessing firms are tied to state farms investments by combining novel historic datasets and map data from the National Archives of Ethiopia and the

Ministry of Agriculture to contemporary commercial farm and manufacturing data from the

Central Statistics Agency. Using plot level commercial farm data, I identify how crop yields drive commercial farm agglomeration patterns. I also empirically examine Ethiopia’s agroprocessing industrialization strategy by discerning whether firms gain productivity advantages from sourcing inputs domestically. Finally, I analyze linkages between the commercial farm and agroprocessing sector by matching county level commercial farm production data to firm level manufacturing data.

I find that state farm investments during the Derg regime altered the geospatial distribution of commercial farm and agroprocessing production, but the underlying farm type matters. Woredas (counties) that had a historic state farm investment between 1980-1990 have at least 124% more commercial farm area and 137% more agroprocessing production, but horticulture state farms (mainly fruits and vegetables) have stronger spatial linkages to the modern-day agroprocessing sector. Commercial farms that grow staple crops have been able to successfully form new commercial farm enclaves in low-lying regions while horticulture farms have been constrained to highland regions.

Furthermore, I assess the underlying mechanisms that impede current commercial farm productivity and find path dependency effects associated with state farms. Commercial farm plots located within staple crop state farm counties are less productive than other commercial farm plots due to inadequate investment in public infrastructure, like irrigation. Commercial farm plots located within horticulture state farm counties are more productive than commercial farms located in other counties. These disparities arise from different location choices and investment decisions made by national planners during the Derg regime. I also find that agroprocessing firms that source inputs domestically are more productive than firms that source inputs abroad.

Yet, the commercial farm sector played a negligible role in feeding the recent expansion of the agroprocessing sector. Although agroprocessing firms located in woredas with commercial farms gain input sourcing advantages, especially for horticulture, these advantages are marginal and do not translate into improved firm level worker productivity.

My research further bolsters findings that commercial farming location criteria is highly selective in developing contexts (Glover and Jones 2016), but private farm location decision making is also largely shaped by government action and constrained by historic incidences (Hall

8 et al. 2015). Furthermore, my findings reinforces the merits of the EPRDF’s strategy to strengthen domestic value chains, but questions their laissez-faire approach with the commercial farm sector; without the government acting as the main coordinating agency for rural development in low-lying regions, new commercial farm enclaves for high value added agriculture that have stronger linkages to the agroprocessing sector will likely not emerge. My research also highlights avenues for further research. The underlying mechanisms driving the spatial relationships between the commercial farm sector and agroprocessing sector are not well understood within the context of second city development. More research should be done to understand how commercial farm investments can spur agroprocessing agglomeration. Since agroprocessing firms refine inputs that are bulky and perishable and thus often locate in the agricultural heartland, their location choices could promote spatial integration.

But, in order to further understand the consequences of state farms on the present-day commercial farming and agroprocessing sector, we must first analyze the political environment in which state farms were formed. Section II of the paper details how state farms were created and privatized alongside various agricultural development strategies across the Imperial (1942-

1974), Derg (1974-1991), and EPRDF (1991-present) regimes. Section III reviews relevant literature and outlines the methodology. Section IV explains the data sources, georeferencing and name matching strategies used to clean the data. Sections V and VI cover the data analysis and conclusion.

II. Historical Background

The theoretical intuition guiding Ethiopia’s development strategies has been well-defined for decades. Economist Albert Hirschman famously posited that forward and backward linkages between agriculture and manufacturing cause spillover effects: productivity growth in the agricultural sector creates cheaper inputs while also driving down labor costs as farmers leave agriculture to find employment in manufacturing (Hirschman 1958). Backward linkages improve agricultural productivity and production since there are relatively less farmers, increasing the marginal product of labor while agroprocessing firms also demand more commodities.

Despite these widely understood effects, there is little consensus on the underlying mechanisms driving this process or concrete steps policy makers can take in order to induce structural transformation through agriculture. This has followed an unsettling history of limited congruity between research, policy advice, and government action. Across three different political regimes- Imperial (1942-1974), Derg (1974-1991), and EPRDF (1991-present)-

Ethiopian policy makers have grappled with this ambiguity and have taken drastically different measures to transform the agricultural sector with mixed results. During the 1960s to 1970s, the

World Bank’s agricultural modernization strategy focused primarily on large-scale commercial farms, yet they later reversed policy and preferred development initiatives supporting small-scale agriculture over efficiency concerns of large-scale farming (Sharp et al. 2007). Recently, Collier and Dercon (2014) questioned the approach many NGOs and multilateral organizations have taken by unilaterally supporting small-scale farmers over other forms of agricultural production.

Considering the divergent experiences between African countries’ agricultural performance and

Asian countries’ “Green Revolution”, African policy makers are left with more questions than answers.

a. Imperial Regime (1942-1974)

Beginning in 1957, the ruling Ethiopian King Haile Selassie set out to transform his country’s economy by following long-term development strategies after consulting technical experts from United States (Stanford 1969). Selassie would create three, “Five Year Plans”, with the last interrupted by the Socialist Revolution in 1974. The First Five Year Plan (FFYP, 1957-

1961) focused on five priority areas: infrastructure development, education, agricultural growth, agroprocessing, and mobilize financial resources (SFYP, 2/7). The government’s efforts were largely successful in building roadways, ports, and expanding electrical capacity. However agricultural investments were generally allocated to small-scale rural investments that failed to address structural problems surrounding Ethiopia’s feudal land tenure system, which critics argue distorted small peasant holders from investing in their land and adopting modern methods.

Ultimately, the government was forced to import grain crops, the staple crop of rural farmers, to meet local consumption demand.

During the Second Five Year Plan (SFYP, 1962-1967), the government focused on creating forward and backward linkages between agroprocessing sectors and large-scale commercial farms noting that, “The agro-industrial complex is an important part of SFYP aimed at utilizing the already existing wealth for development purposes, at promoting exports and increasing the domestic saving capacity” (SFYP, 4/2). Commercial farms would provide the necessary inputs for agroprocessing firms to refine and then export processed goods abroad.

Foreign exchange earnings could be reinvested into other priority manufacturing sectors, mainly chemicals and steel. Because these sectors require large foreign exchange surpluses to purchase capital and inputs from abroad, the government wanted to lay the groundwork for industrial diversification by creating a reliable flow of foreign exchange.

Following the World Bank’s agricultural modernization strategy, large-scale commercial farms were recommended, and Ethiopia’s agricultural policies gave preferential treatment to export-oriented commercial farming (Dessalegn et al. 2006). To kickstart commercial farming, the government gave special privileges to investors, including five-year tax exemptions and additional benefits for select crops. The two most important investors would be the Dutch HVA company for sugar and the British Mitchell Cotts company for cotton, which would occupy thousands of hectares of land in the Shoa and Sidamo region (Clapham 2019, 35). Although the government sought to spend 53% of its agricultural budget to support large scale farms, they later postponed their efforts due to investment constraints. Instead, they primarily invested in agroprocessing sectors and institutional reforms in the agricultural sector, such as agricultural extension services and research, to ensure they could support commercial farm projects in the next development plan.

During the Third Five Year Plan (TFYP, 1968-1973), the Imperial regime finally managed to attract large-scale foreign and domestic investment into commercial farming. They rolled over the same investment benefits from the SFYP and opened new areas in the north- western lowlands and Awash Valley for commercial agriculture. Commercial farms were initially located in rural areas where cropland was available and suitable for large-scale production. Large-scale commercial agriculture also offered seasonal employment opportunities to local farmers. This, however, exacerbated complicated land tenure relationships (Dessalegn et al. 2006). By 1973, there were over 1,450 commercial farms operating in Ethiopia (Directory of

Agriculture 1973).

Although the economy grew on average 3-4% during this fifteen-year period, the government’s top-down development approach had mixed results. The manufacturing sector

12 grew annually by 10%, but the Imperial regime failed to transform the country from an agrarian to industrial economy as 90% of workers were still employed in the agricultural sector. Peasant incomes did not grow since the government underinvested in peasant agriculture and did not address its structural problems, such as land tenure and poor education. The Imperial regime’s development strategy was also inherently political as it supported the urban elites concentrated in

Addis Ababa and directed the development of infrastructure projects that only exacerbated regional inequalities (Griffen and Hay 1985). Peasant agricultural production was only favorable in a few areas, and meanwhile most peasants remained disconnected from markets and used the same farming techniques and inputs they used for centuries. Historians argue the Imperial regime’s development strategy only further entrenched the political establishment in Addis

Ababa, and did not resolve the inherent contradiction of Ethiopia’s political economy, but rather reinforced, “the deeply unequal relationship between those who controlled the state and the areas of the countryside in which this development process necessarily took place” (Clapham 2019,

35).

Map 1: Woredas with Commercial Farms in 1973 vs. Woredas with State Farms in 1985

Sources: 1973 Directory of Agriculture, 1985-1987 Regional Atlases of Ethiopia

b. Derg Regime (1974-1991)

Once the Communist Derg regime overthrew the Imperial regime in 1974, they upended the preexisting feudal system in hopes of fulfilling its promise to achieve state sponsored development. On March 4th, 1975, the Derg nationalized all land holdings larger than 10 hectares and unshackled tenants from their landlords by redistributing land to peasant holder farmers

(PMAC 1975). State farms were initially not a priority among national planners and only comprised 0.6% of all cultivated land in Ethiopia in 1979 (Table 1). Government policy was focused on realigning tenure systems, implementing price controls, and creating market boards.

The pre-existing land tenure system during the Imperial regime preserved basic cultivation practices since a small group of landlords extracted a surplus over tenants. Nonexistent property rights reduced incentives for peasants to invest in their land (Belete et al. 1991). Peasants were grouped into Peasant Associations that comprised 300-400 farmers and approximately 800 hectares. For the first time, peasants had access to their own land and were no longer required to pay up to 75% of their harvest to landlords. Since peasant holders now had stronger land rights, national planners believed changes in rural holding incentive structures would ultimately lead to massive improvements in agricultural productivity. By 1985, there were over 5.6 million households within Peasant Associations and they comprised the vast majority of agricultural output (Table 1). Renting land was abolished nor could farmers employ workers seasonally

(Sharp et al. 2007). Peasant holding agriculture was strictly confined to the family unit. Although some peasants formed producer cooperatives, they only represented 1-2% of area and output in the agricultural sector.

Private commercial farmland was nationalized, and most commercial farms were transformed into large-scale state farms by combining nearby commercial farms into mega-

14 farms. In certain woredas, commercial farms were converted into cooperative farms or turned over to Peasant Associations. Counties that had smaller populations and less land under cultivation were more likely to gain a state farm since it was easier to expand state farm cultivated land without interfering with peasant land holdings. Only small horticulture farms located in the highlands regions due to high population density. For instance, in 1990 over 88% of peasant holding farmers were in the highlands region, which only comprises 40% of

Ethiopia’s landmass (Belete et al. 1991). Commercial farms that were difficult to redistribute to peasant farmers were more likely converted into state farms (Griffen and Hay 1985). Yet, from

1974-1979, agricultural productivity declined 1.4% and with a population growth rate of 3%, national planners were forced to reformulate their agricultural policy in order to meet Ethiopia’s growing food demands (Griffen and Hay 1985).

Table 1: Staple Crop Production by Agricultural Sector during the Derg Regime

Year Peasant Farmers Co-Operatives State Farms Area (%) Output (%) Area Output Area Output

1975/76 98.4 97.9 1.0 0.8 0.6 1.3

1976/77 98.9 98.8 0.8 0.5 0.3 0.7

1977/78 98.9 98.0 0.8 0.6 0.3 1.4

1978/79 98.7 97.5 0.7 0.7 0.6 1.8

1979/80 96.0 96.1 2.4 1.5 1.6 2.4

1980/81 95.9 95.1 1.5 1.0 2.6 3.9

1981/82 94.3 94.2 2.2 1.2 3.5 4.6

1982/83 95.1 95.3 1.9 1.3 3.0 3.4

Source: Ghose 1985, p.132

To address endemic food shortages, the Ministry of State Farms Development (MSFD) was created on May 2nd, 1979 with four explicit mandates: 1.) Organize state farms that specialize in cereals, livestock, fisheries, and fruit and vegetable production 2.) Establish model state farms that will encourage uptake of modern farming techniques 3.) Produce commodities for domestic consumption and export markets 4.) Produce raw materials for domestic processing and agroprocessing industries (Mirotchie and Taylor 1993, 188). MSFD moved quickly to consolidate tractable land for large-scale commercial farm production and within a year the total number of hectares (ha) increased from 64,000 ha. to 293,000 ha. By 1985, there were 54 state farms in operation. Although most farms grew staple crops, horticulture farms were also later developed. Staple crop state farms were much larger than horticulture state farms and more capital intensive, employing less workers per hectare (Table 2). To put the size of these farms into perspective, Peasant Associations had 300-400 family units for 800 hectares, whereas staple crop state farms only employed 200-400 permanent workers on 5,000+ hectares.

Table 2: Employment Intensity by Select State Farms

State Farm Main Crop Area Permanent Workers Hectares per (ha.) Worker

Dubti Cotton 5,266 291 18.10

Dit Bahari Cotton 5,347 406 13.17

Asaita Cotton 1,917 103 18.61

Mille Cotton 875 159 5.50

Melkasedi Banana 700 992 0.71

Amibara Angele Cotton 2,652 359 7.39

A.I.P. Cotton 1,805 165 10.94

Defan Belhame Cotton 1,400 89 15.73

Gewane Cotton 2,083 626 3.33

Adelle Wheat 6,000 149 40.27

Ardaita Wheat 7,200 196 36.73

Dikisis Wheat 10,320 252 40.95

Garadella Wheat 11,350 294 38.61

Gofer Wheat 8,500 220 38.64

Dinkite Wheat 10,000 241 41.49

Source: Griffen and Hay 1985, p. 55

Upon inception, state farms immediately had efficiency concerns and ran up large deficits. State farms required investments in rural road networks, improved roads to access regional markets, irrigation schemes, machinery, inputs, and semiskilled and skilled laborers to manage the farms. When the Derg nationalized commercial farms in 1975, commercial farms had only begun to operate so large-scale infrastructure and capital investments were needed to combine commercial farmland into large state farms. State farms also consumed 75% of improved seeds and absorbed 12-15% of the total agricultural budget, while at peak only

17 produced 4-6% of total agricultural production (World Bank 1987). Between 1981-1985, only five state farms had a single year of positive profits (MSFD Financial Performance 1987).

Increased investment in the state farm sector did not lead to productivity gains. Following a translog production function, Mirotchie and Taylor (1993) found that staple crop state farms exhibited constant returns to scale. Staple crop state farms were overcapitalized and underutilized labor. The elasticity of substitution between labor and overutilized inputs was low, suggesting staple crop state farms were endemically flawed. Farm managers could not simply substitute labor for capital in hopes of improving agricultural productivity (Mirotchie and Taylor

1993). State farms also faced technical and administrative challenges as, “few of the farm managers have experience in running large agricultural enterprises and many are not properly trained” (Griffen and Hay 1985, 55-56). Effective commercial farming requires complementary skills that are difficult to simply learn on the job such as managerial skills, handling knowledge diffusion, managing risk, and science understanding (Collier and Dercon 2014, 5). Since management was poorly trained, they were ill-equipped to oversee the expansion of state farms.

State farms also had trouble employing seasonal workers. Since state farms required workers when peasants were tending to their own farms, rural labor market supply was limited, especially since many state farms were in unpopulated middle and low-lying regions. Managers could not pay more than the legal minimum wage of 1.92 birr per day even though the market rate for seasonal labor was at least twice the minimum wage (Griffen and Hay 1985, 54), forcing farms to bus in seasonal workers from urban regions. For instance, two sesame producing state farms in the modern day Kafta-Humera woreda required bussing 50,000 seasonal workers during harvest, requiring them to not only pay their wages, but also provide food and shelter (MSFD

Main Report 1986). This not only was costly, but also reduced the potential for knowledge

18 spillover onto local peasant holding farmers. The state farm development strategy ultimately drained resources from Peasant Associations and co-operative farms by restricting investment in rural infrastructure and modern inputs for peasant farmers, coupled with agricultural marketing boards that depressed food prices. Although there were some examples of government investment in agroprocessing sectors, there was neither the necessary labor agricultural productivity growth to push workers out of the agricultural sector nor strong productivity gains in the state-sponsored industrial sector to support structural change (Shiferaw et al. 2019, 142).

Furthermore, the 1984-1985 famine became emblematic of the government’s inadequacy to deliver on their central political goal of agricultural growth through rural land reform. This further compounded the civil war in the northern region as people in Tigray and Eritrea viewed the Derg incapable of governing (Clapham 2019, 35). The Derg regime ultimately failed to transform the economy and by the 1990s, Ethiopian real agricultural productivity was 55% of its

1960s levels (Adamopolous 2018).

Chart 1: Economic Growth in Ethiopia by Sector (1975-2017)

Source: Shiferaw and Manyazewal 2019

c. EPRDF (1991-present)

Once the transitional government took control after the civil war ended, they immediately partnered with donor countries and multi-lateral organizations. The EPRDF and World Bank coordinated the Emergency Recovery and Reconstruction Program to rehabilitate infrastructure and implement liberalizing reforms to achieve macroeconomic stability (Shiferaw et al. 2019,

143). As part of its mandate, the EPRDF orchestrated a radically different long-term development strategy that prioritized peasant holder agriculture over large-scale state farms.

Agricultural Development Led Industrialization (ADLI) would guide the overarching development agenda for the next 10-15 years by providing modern inputs, extension services, and rural road networks to smallholders with the hopes of creating forward linkages to agroprocessing firms. This development agenda aligned with the “agricultural-first” narrative wherein structural transformation could only occur by first sustaining agricultural productivity growth. This closed economy theory presumes that as productivity increases, consumer consumption bundles would change and begin demanding manufactured goods while also mobilizing the necessary capital investment in the manufacturing sector. ADLI’s message was also consistent with its main rural support base in Tigray, making it politically expedient.

Although ADLI rapidly increased agricultural production with year-over-year growth averaging approximately 9% between 1995-2010, most of the gains were associated with expanding land cultivation area. After 2010, a large portion of cereal production growth was associated with productivity growth (Dercon et al. 2019, 459). Road expansion between 1996-

2010 promoted spatial integration for both the peasant holding and manufacturing sector. The

Road Sector Development Program (RSDP) invested US $8.07 billion into Ethiopian roads and expanded roadway networks from 26,550 to 53,997 km. By reducing transportation costs

20 associated with intermediate inputs and expanding market access, the RSDP increased average staple yields by 13.6%, accounting for 10% of the overall production growth (Adamapolous

2018). Shiferaw et al. (2015) found increased road networks reduced the spatial concentration of firms. ADLI’s success has largely been attributed to supporting agricultural growth, but before

2010, agricultural growth did not spill-over onto manufacturing. Linkages between smallholder agriculture and agroprocessing did not materialize and by 2010 manufacturing only comprised

5% of Ethiopia’s GDP (Shiferaw et al. 2019, 138).

Beginning in 2010, the EPRDF readjusted their development strategy and focused on inducing structural transformation through industrial policy with the first Growth and

Transformation Plan (GTP I, 2010-2015) and GTP II (2015-2010). The National Planning

Commission and the Ethiopian Investment Commission have been the primary agencies tasked to construct industrial parks across the country and finance foreign manufacturer relocation with hopes of spurring manufacturing agglomeration. Policy makers are still focused on creating backward linkages to the agricultural sector and create multifaceted strategies that can promote export oriented manufacturing while also support a revolution within agriculture (Dercon et al.

2019, 449). The Ethiopian government is nearing the completion of four agroprocessing industrial parks (IAIP) located in Central Eastern Oromia, South West Amhara, Western Tigray and Eastern SNNP (Map 2).

Map 2: Regional Location of IAIPs and Catchment Areas (UNIDO 2018)

Agroprocessing industrial parks provide the necessary infrastructure, including energy, water, road network, and telecommunication, to induce foreign and domestic agroprocessing firms to relocate near their input supply chains. Furthermore, they provide a steady supply of cheap labor to firms by investing in housing and schools for workers’ families. By grouping foreign and domestic firms together in the same park, the government hopes agglomeration benefits will also arise from knowledge spill over. The IAIP’s strategy relies solely on inputs being supplied by peasant holding agriculture. The government hopes to meet agroprocessing input demand by quickly transforming the peasant holding agriculture sector. Rural transformation centers (RCTs) have been implemented throughout the regional catchment areas and provide warehouses, input supply, sorting, grading, and extension services to the smallholder farms with the hopes of linking farmers to commercial value chains (UNIDO 2018, 7).

Chart 3: Peasant Holder Agricultural Plot and Labor Productivity (Taffesse 2019, 474)

This strategy capitalizes on the size and scope of the peasant holding farm sector, which comprises 95% of agricultural output in Ethiopia, as well as recent history. From 2003-2014, peasant holder grain productivity growth rates averaged 5.6% (Taffesse 2019, 476). Productivity improvements continued after 2010 and were attributed to the expansion of extension services that increased farmers’ access to better farming practices, increasing education levels, improved land tenure rights, and the adoption of improved seeds and fertilizer (Taffesse 2019 and Bachewe et al. 2018). For instance, chemical fertilizer application increased 8.5% annually and by 2015,

50% of grain farmland used fertilizers. By 2014, 10% of grain area was planted with improved seeds. This was only possible with the expansion of extension service coverage to encourage farmers to use modern inputs (Taffesse 2019).

After 2010, growth in agricultural output and productivity catalyzed value added growth in the agroprocessing sector. From 2012-2015, food and beverage value added increased from

5,570,603 Ethiopian birr to 21,257,826 birr (CSA LLMIS Report). Most of the agroprocessing value added growth during this time period was associated with firms that refined staple crops.

By 2016, 61.3% of the value addition in the food and beverage processing sector was attributed

23 to agroprocessing firms that refined staple crops as inputs. From 2012-2015, the number of malt liquor establishments rose from 9 to 18, the number of bakery firms increased from 210 to 272 and the number of grain mills increased from 216 to 344 firms. New firms formed to take advantage of the growth in output.

The continued application of modern inputs, coupled with expanding extension services, will undoubtedly continue to improve peasant holding land productivity. But, there are growing concerns, given the small-scale of peasant holding agriculture, that potential output growth is nearing its limits, “A large proportion of them [plots] appear to be too small to produce significantly higher levels of output under the existing crop mix and farming technology. In part reflecting these constrained opportunities, farm households seek alternative income-generating activities” (Taffesse 2019, 480). These concerns may be justified since value added growth for agroprocessing firms has recently abated, decreasing 9.4% between 2016-2017 (CSA LLMIS

Report 2016-2017). Between 2015-2016, the number of grain mills fell from 344 to 287 while the number of bakeries fell from 272 to 229 and the number of malt liquor companies fell from

18 to 16 (CSA LLMIS Report 2015-2016). As land productivity improves and agricultural labor productivity continues to fall, peasant holding farmers will search for alternative off-farm activities and may begin to migrate to urban areas at faster rates. But this will only be possible if there are corresponding manufacturing jobs.

Although commercial farms do not play a role within the agroprocessing industrial park strategy, they provide a disproportionate amount raw inputs to the agroprocessing sector. In

2011, commercial farm production only amounted to 5% of total agricultural production, but they provide 25% of the raw input supply to domestic agroprocessing firms (UNDP 2011). The commercial farm sector production is centered around domestic markets, with only 25% of

24 production intended for export markets. By 2011, there were over 10 million hectares of commercial farmland leased, far larger than the total accumulation of state farmland during the

Derg period, which amounted to 4.6 million hectares (UNDP 2011 and Table 1).

Beginning in 2009, the government took control of commercial farmland leasing from regional governments and required that all regions identify uncultivated land that could be leased to private commercial farms (Lavers et al. 2012). Beforehand, the private farmland was leased by regional governments and the federal government would only encourage commercial farm development by privatizing state farmland. Beginning in 1996, the EPRDF began state farm privatization and by 2016 they had privatized over 26 state farms (EPA 2018). Now, the federal government policy promotes commercial farmland investment in low-lying regions in order to spur commercial farm agglomeration in new areas and encourage rural development within pastoral communities. The government gave long term leases, five-year tax holidays, and cheap loans from the Development Bank of Ethiopia to incentivize investment within these regions.

Cotula et al. (2014) found that three low-lying regions, Benishangul-Gumuz, Gambella, and

South Nations, Nationalities and Peoples’ Region account for 80% of recently leased commercial farmland. These projects have been well-documented among news outlets and critics have attested these leases incentivize land-grabbing by stripping land away from smallholders and pastoralists, while also promoting speculative practices (Hall et al. 2015 and Burgis 2016).

As well, the government has recently coordinated state projects to replicate the success of sugar estates constructed by the Dutch HVA company during the Imperial regime. Beginning in

2010 under GTP-I, the Ethiopian Sugar Corporation (ESC) directed investment to existing sugar plantations to expand production and was also in charge of constructing the Kuraz sugar plantation. The 175,000 hectare plantation located in the SNNP region was supposed to increase

25 domestic processing capacity by 700% and increase production by 90% (Kamski 2016). Instead, by 2016 only 10,000 hectares have been cultivated and the project has been marred by inadequate financing, technical plant design flaws, costly dam and irrigation schemes, and environmental degradation that negatively affects local communities (Kamski 2018). Critics have argued this top-down approach is another “white-elephant” project that will fail to meet domestic sugar demand and not provide enough jobs to promote urbanization and structural change in a remote section of the country (Street 2020).

As of 2020, the Kuraz Sugar Development project continues, but for the first time, government officials are looking at complementary reform measures to support productivity growth within the pre-existing commercial farm sector. The newly released “Homegrown

Economic Reform Agenda” in September 2019 explicitly outlines sectoral strategies to liberalize financial markets, privatize state owned enterprises, and invest in information-technology infrastructure to spur foreign direct investment. The agricultural sector reforms consist of creating a legal framework to allow peasant holding farms to lease land to commercial farms, encourage private sector investment in agricultural R&D, and explore public-private partnerships to expand medium and large-scale irrigation infrastructure (Office of the Prime Minister 2019).

In light of this new policy agenda, there is renewed interest in understanding the production structure of the commercial farm sector and specific policy measures that can enhance commercial farm productivity growth.

III. Conceptual Framework & Methodology My analysis follows a dual-pronged approach that analyzes the current and historic determinants of commercial farm agglomeration and its corresponding effect on the agroprocessing sector. Due to the size and scope of commercial farm nationalization and state farm conversion, state farm investments likely had path-dependent effects on these sectors. But the spatial consequences of state farm investments on current agglomeration patterns are not well understood. Furthermore, it remains unclear how recent foreign direct investment (FDI) has either reinforced commercial agglomeration or promoted spatial diffusion by forming new commercial farm enclaves. Finally, more research must be conducted to understand whether the government should prioritize investing in domestic value chains and whether recent commercial farm growth contributed to agroprocessing value added growth. I begin by exploring the potential underlying mechanisms that altered commercial farmland production and productivity outcomes. I then explain the empirical strategy used to weed out the endogeneity of state farm site selection and then transition towards the identification strategy used to assess forward linkages to the agroprocessing sector. a. Commercial Farm Production

There is a growing body of literature detailing how historical incidences shaped geospatial inequality patterns on the African continent through public sector infrastructure investment. Roessler et al. (2019) found that present night-time lights are highly correlated with colonial era cash crop production since colonial governments invested exhaustively in public goods necessary to export cash crops abroad. These sunk cost investments could not be reallocated, and instead economic activity conglomerated around these enclaves, reshaping

Africa’s economic geography. A more recent example of this phenomenon is Liberian natural

27 resource FDI during the Johnson-Sirleaf administration. Bunte et al. (2018) found Liberian FDI promoted agglomeration within rural areas since projects were only approved if they provided public goods to specific geographic areas. Not only do state farms likely affect present commercial farm locations directly through state farm privatization into commercial farm schemes, but there are likely agglomeration benefits if state farms invested in rural road networks.

Glover and Jones (2016) found that commercial farming FDI in Mozambique is highly selective, preferring to locate near existing infrastructure and local markets. The relationship between commercial farm location, infrastructure, and market access may be further pronounced in Ethiopia because until recently, the country had underinvested in rural road networks and irrigation schemes. The public resources needed to facilitate the connection to markets, stabilize input supplies, and improve crop yields are relatively scarce in comparison to other African countries. Furthermore, commercial farm agglomeration could reduce future investment costs through joint-financing and incentivizes public sector investment in remote areas to take advantage of the positive externalities. Due to Ethiopia’s unique history, there is little research conducted on ulterior pathways in which public sector investment in agricultural services influences commercial farm agglomeration. Woredas with state farm woredas may have also reduced the barriers to entry for commercial farms through preexisting modern input supply chains, extension services, and storage facilities that were intended for state farms.

Although pre-existing infrastructure and agro-climatic suitability likely influence the location decision making of commercial farms, private location choices are only made possible with state action. In a continent-wide comparative analysis of commercial farm land-grabbing,

Hall et al. (2015) finds that large-scale transactions are directed by federal bureaucracies. African

28 states often overstep their jurisdiction on local municipalities to ensure commercial farms have cultivated land, and in some cases, governments intimidate smallholders to accept land-leasing agreements. Commercial farm location choices are confined to potential areas the government can “open up” for investment. Hall et al. (2015) also finds these location choices are inextricably linked to water access for irrigation schemes and historic determinants. To mitigate the removal and relocation of smallholder farmers, governments will try to privatize uncultivated land area, like customary lands, or state farm and colonial plantations that already have the structure in place for large-scale commercial farming.

Ethiopia is no exception to these continental trends as they have already privatized 26 state farms, while also supporting investment projects within low-lying regions that are relatively unpopulated. Lavers et al. (2016) found that Ethiopian government led commercial farm investments decisions were driven by four key trends: 1.) attempts to implement private contract- farming schemes in the highland regions 2.) leasing communal lands, 10-20 ha., to domestic investors to promote commercial cultivation in the highlands 3.) state farm privatization with the hopes that private investment can address long-standing productivity concerns 4.) vast expansion of commercial farm cultivation for low-lying regions. For instance, land targeted for future commercial farm investment represents 42% of Gambella’s land area, one of the low-lying regions in Ethiopia. These studies highlight the rationale behind where the EPRDF has tried to locate commercial farm investment. But there has been no research conducted to understand whether these governments’ initiatives have been effective at spurring commercial farm agglomeration in new areas.

(1) ln(Y)wt = β0 + β1Dw + Xwt + εwt

To investigate the overall impact state farm investments have on the geospatial distribution of commercial farm production, equation (1) follows a simple Pooled OLS approach where ln(Y)wt is the logged total production in quintals of woreda w at time t where t ϵ (2011,

2013, 2014, 2015). Xwt is a vector of woreda level geographic, road density, and population density variables to control for any additional characteristics that may affect commercial farm production and 휀푤푡 is the residual error. The model also includes regional and year fixed effects and the models were also run using random effects. Dw is a dummy variable that takes the value of “1” if a woreda had a state farm at any point between 1980-1990, otherwise Dw takes the value of “0”. Since we expect state farms altered the geospatial distribution of present commercial farm activity, we expect β1 to have a positive coefficient. It is important that we control for woreda level geographic characteristics since agro-climatic suitability also drives current commercial farm location choices and production. As well, it is important to include time fixed effects since they capture time varying macroeconomic shocks, while region fixed effects control for long run political factors that influence commercial farm location and production, which may be heterogenous across regions. b. Commercial Farm Productivity

Current productivity determinants of commercial farm plots in Ethiopia are partly driven by the production structure of commercial farms. Ethiopian commercial farms can be categorized into two groups: well-established farms and speculators (UNDP 2011). Well-established farms produce commodities at high productivity levels and pilot-test their own field crops. Speculators adopt low cost, low output models that maximize short term profit with minimal fixed

30 investment expenditure and this group comprised 50% of all commercial farms. Speculators may be more inclined to locate in low-lying regions since they can boost output by increasing their cultivated land area instead of investing in modern inputs that would improve productivity. Low- lying regions may also have lower plot productivity levels since they have restricted access to modern input markets. There are also shortages in modern inputs since domestic seed corporations are unable to supply enough improved seeds to the commercial farm sector (Mbata

2012). Commercial farms often resort to importing inputs, but this may be too costly for commercial farms located in low-lying regions far-away from Addis Ababa.

On the other hand, there has been no research conducted to understand the historic determinants of plot level commercial farm productivity in Ethiopia. The hypothesized relationship between present commercial farmland productivity and historic state farm development likely depends on farm-type. Mirotchie and Taylor (1993) found that staple crop state farms were overcapitalized, and their large-scale production structure only saw constant returns to scale. On the other hand, horticulture state farms were smaller, more labor intensive, produced high value added cash crops that required irrigation, and were located much closer to towns (MSFD Horticulture Report 1987-1989). These different location choices and modern input access caused large productivity differentials between staple crop and horticulture state farms during the Derg regime (Gabriel 1990 and MSFD Horticulture Report 1987-1989).

These productivity effects may persist since there have been no attempts made to reform large-scale commercial farms in Ethiopia. The government hoped private sector investment within preexisting state farms would lead to productivity convergence (Lavers et al. 2016). But privatization likely did not address these problems since Tefera & Lu (2018) found that only

67.4% of private commercial farmland was cultivated from 1995-2016, as much of the land was

31 left unused. The operational structure of staple crop state farms likely exists today as many commercial farm investors have failed to invest in the necessary rural infrastructure and modern inputs to cultivate all their government allotted commercial farmland since it is likely more lucrative to adopt low cost, low output production models.

Y (2) ln( ) = β + β D + B S + B I + X + ε A pwt 0 1 w 2 pwt 3 pwt wt pwt

To assess plot level productivity, equation (2) follows a similar OLS approach as equation (1), but instead we link woreda characteristics Xwt and the state farm dummy variable

Dw to each commercial farm plot p at time t. Furthermore, in order to assess additional mechanisms that affect current crop productivity, equation (2) includes modern input data at the plot level, Spwt and Ipwt, which are dummy variables denoting whether a plot uses improved- seeds or has access to irrigation, respectively. We expect these dummy variables to also have positive coefficients since modern inputs improve plot productivity.

However, OLS estimates are likely biased since they do not control for state farm woreda selection. It is unrealistic to simply compare all woredas that previously had a state farm to woredas that did not have a state farm because woreda characteristics can change over time, even after controlling for modern geographic, socioeconomic, and demographic differences.

Furthermore, due to the size and scale of state farm investments, the selection criteria national planners used during the Derg may be different than current private investors. In order to mitigate the selection bias effect, we must restrict the sample space of control participants within the analysis by using a spatial propensity score matching technique.

32 c. Spatial Propensity Score Matching

Let Y0 and Y1 be the outcome of interest for two separate woredas and X be a vector of observable characteristics. The starting assumption for the propensity score matching technique formulated by Rosenbaum and Rubin (1983) is that treatment selection, denoted by D ∈ (0,1), is independent from other participants, or woredas in this scenario (3). We can eliminate the selection bias by subtracting the means of participants selected for treatment, state farm woreda selection, against participants with similar characteristics but did not receive treatment (4).

3 (Y0 , Y1 ) ⊥ D | X

4 E[(Y0 , Y1 )| D=1, P(X)] = E[(Y1 )| D=1, P(X)] - E[(Y0)| D=0, P(X)]

5 B(X) = E[(Y0 )| D=1, P(X)] - E[(Y0)| D=0, P(X)]=0 Yet, equation (4) is only true if and only if equation (5) equals 0 where B(X) is the measured selection bias. Bias is estimated by taking the difference between the hypothetical outcomes of control participants if they had received treatment and their actual outcomes (Heckman et al.

1998). However, bias can also arise from a lack of common support or unobservable errors correlated with the outcome variable due to an incomplete set of observables. Propensity score matching techniques match on the probability of receiving treatment, equation (6), while conditional matching approaches simply match directly on the observable characteristics, X.

6 P(X)= Pr (D=1 | Y0 , Y1 , X)= Pr (D=1 | X)

7 Cor(X D) ≠ 0 & Cor(X,Y)=0

8 Sup(| D=1) ∩ Sup(X| D=0)≠ ∅

Propensity matching is more effective than conditional matching since observable characteristics are correlated with site selection, D, but may not be directly correlated with the outcome variable (7). This set up is conceptually similar to an instrumental variable approach; we do not include the instrumental variable alongside the endogenous variable in the final regression, but rather run a preliminary regression to weed out endogeneity. The inclusion of characteristic variables that influence site selection in the outcome regression would be inadequate. A more effective strategy is matching each program participant with a counterfactual participant that has similar observable characteristics, and then restrict the analysis to this segment. Optimal matching techniques use the coefficients generated from a logit model to calculate the individual participant probabilities of receiving treatment. Counterfactual matches between participants in the control and treatment are made by searching across the entire control group population and finding the control participants that minimizes the total propensity distance. The propensity distance is measured by subtracting the propensity score values of treated participants versus control participants.

Propensity score matching techniques only work if 0 < P(X) < 1 and there exists a common support between the treatment and control groups (8). These conditions highlight the perplexing irony of this approach; the logit function must control for all observable characteristics that influence selection, but the function cannot perfectly predict control and treatment participants, otherwise P(X)=1 or P(X)=0, and thereby violating one of the necessary assumptions of the model. After controlling for all observable data, variation should still exist, meaning participants with the same attributes are randomly sorted into control and treatment groups.

Yet, what if the initial assumption, equation (3), is not held and treatment selection is not independently made across participants? Suppose program participation is spatially correlated, meaning site selection of state farms is not only contingent upon the woreda characteristics, but also on the location of other selected state farm woredas. Since program participation is no longer an independent decision, the propensity score matching technique’s model errors would be correlated, Cor(휀1, 휀 0 ) ≠ 0. Chagas et al. (2011) presents an ulterior propensity matching estimation strategy to deal with spatial correlation by estimating Bayesian spatial error and spatial lag logit models to construct propensity matches. I propose a simpler estimation strategy where we include a characteristic variable in the logit model that captures the spatial relationship of the data generating process. This is a viable estimation strategy because the model errors would no longer be spatially correlated since we control for the effect that program selection has on the outcomes other participants. Yet, this is only effective if the characteristic variable captures the spatial relationship between program participants.

Although we cannot explain what underlying mechanism causes the spatial relationship, this is not a relevant objection. For instance, reduced form models distill the complexity of structural models into an estimated model. If propensity estimation equation includes a variable that captures the spatial relationship and all other characteristic variables relevant to selection are included, then the propensity score matching model is non-biased, estimated by equation (9). The average treatment effect is found by subtracting each treatment participant Y1i indexed by iϵI1 against each propensity-matched control participant Y0i indexed jϵI0. Each control participants value is weighted by a positive coefficient matrix Wij that gives observations that have smaller propensity distances more weight.

1 9 E[(Y1 – Y0 )| D=1, P(X)] = = ∑ [ Y - ∑ W Y ] N iϵI1 1i iϵI0 ij 0j

′ ′ ′ 10 ln(Ywt − Y′wt) = (β0−β′0) + β1(Dw − Dw) + (Xwt − X wt) + (εpwt − ε pwt)

′ Y Y ′ ′ 11 ln( − ) = (β0−β′0) + β1(Dw − Dw) + B2(Spwt − S pwt) + B3(Ipwt − Apwt A pwt

′ ′ ′ I pwt) + (Xwt − X wt) + (εpwt − ε pwt)

The estimated propensity equations (10) and (11) are augmented OLS equations that subtract off the mean of each variable from the counterfactual propensity control group, denoted by the prime. Instead of using the reweighting matrix, W, all woredas are weighted the same and then we restrict the propensity distance for the sensitivity analysis to throw out outlier matches.

This strategy is only effective if all observable criteria that influences site selection is controlled for in the propensity model. Furthermore, even after controlling for all observable criteria there should still exist a common support between the treatment and control groups, suggesting that there is random variation within the selection process.

d. State Farm Location Decision Making Criteria The observable characteristic criteria used for the propensity matching technique can be found in Table 3 and differs for horticulture state farms and staple crop state farms. Selection criteria used by national planners was likely different for horticulture and staple crop farms because not only did they grow different crops, but horticulture state farms generally served export markets and were more labor intensive whereas staple crop state farms were highly mechanized, required a lot more land, and produced crops for domestic consumption.

Throughout my research at the National Archives of Ethiopia and the Ministry of Agriculture, I did not find documents outlining all the explicit criterion MSFD planners used to choose state farm locations. The only criteria that was specifically referenced in historic sources was the

36 presence of pre-existing commercial farmland during the Imperial regime (National Atlas of

Ethiopia 1981). This is understandable since MSFD planners combined pre-existing commercial farm plots together to create state farms.

Yet, other factors clearly played a role in determining state farm woreda selection. Due to the oncoming food shortage in 1979/80 crisis, national planners were forced to make top-down decisions on state farm construction as quickly as possible and expanded state farm area by

166% in less than a year (Table 2). This required state farms to locate in unpopulated areas so they could quickly expand preexisting commercial farmland with minimal displacement of peasant holding farmers. For instance, staple crop state farm counties have population densities that are 3.4 times less than the average Ethiopian county in 2015. Horticulture state farm counties are 2.6 times less densely populated than the average Ethiopian county.

Although I do not find population density data at the woreda level during this time period, other geographic variable correlate highly with historic population density, such as middle-lying (1000-2000m above sea level), low-lying regions (0-1000m above sea level), and distance to historic manufacturing towns. These regions generally had low population densities due to the presence of malaria and other tropical diseases. Other factors that potentially influenced site selection include access to rural road networks, irrigation potential, distance to the capital, mineral deposits, soil types, public expenditures, cattle distribution centers and location of pre-existing major agricultural production regions for small holders.

It is also highly likely that state farm selection was spatially correlated (Map 1) since state farm woredas are spatially grouped together. Even after controlling for woreda geographic characteristics, there were likely benefits for locating state farms nearby. For instance, the

Ministry of State Farms Development was subdivided regionally and there might have been

37 greater returns to scale if state farms were grouped together by reducing coordination costs and sharing a rural labor supply of skilled and semiskilled farm managers. In order to capture this spatial dynamic, a dummy variable denoting whether a county borders a state farm county was included. If spatial correlation exists at the county level, then this dummy variable would effectively capture it since counterfactual counties located next to state farm woredas would have higher propensity scores. To start, both models were run with all potentially relevant variables.

Then, a forward stepwise regression was conducted to find the variables that minimized the

Akaike’s Information Criteriaon (AIC) for both models. Certain variables were thrown out to mitigate overfitting and other variables were included that are theoretically relevant to site selection. To ensure counterfactual matches were not redundant across both models, the large- scale matching function was run and then all counterfactual matches were thrown out before running the horticulture state farm match function.

Table 3: Spatial Propensity Score Matching Logit Models

(1) (2) VARIABLES Horticulture Staple Crop

Number of Commercial 0.0114 0.0182*** Farms in 1973 (0.0220) (0.00686)

Border State Farm County 1.197* 2.195*** (0.687) (0.491)

Distance to Capital (km) -0.0104*** -0.00197 (0.00312) (0.00172)

Road Density (km/km^2) 0.384** (0.170)

River Density (km/km^2) 0.407 (0.282)

Elevation above 2000m 1.302** Dummy (0.595)

Elevation 0-1000m Dummy 0.989* (0.525)

Number of Secondary Schools 0.745** In 1978 (0.317)

Coffee Intensity Dummy 1.413* (0.731)

Cotton Intensity Dummy 1.459* (0.818)

Soil FE Yes Yes

Constant -1.658 -4.469*** (1.187) (0.657)

Observations 684 684 Pseudo R^2 .4035 .2545 Note: Soil FE variables differ for Horticulture and Staple Crop state farms. Road and River Density are logged. Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

Clearly, the selection criteria were different. Horticulture farms were more likely to locate near the capital, Addis Ababa, evidenced by the negative and significant coefficient on the distance to capital variable. As well, Addis Ababa is in the highlands region and the model has a positive coefficient on the 2000m elevation dummy variable. Horticulture state farm counties were more likely to locate near cities, indicated by the positive and significant coefficients on road density and secondary schools. Finally, horticulture farms located near preexisting peasant holding cash crop areas and regions suitable for river irrigation. Staple crop state farms were in remote areas with low population densities since the 0-1000m elevation dummy variable is positive and significant. Public infrastructure variables were not useful in explaining location selection for staple crop state farms and the number of commercial farms in 1973 had a much stronger effect explaining staple crop state farm location. Finally, woreda spatial correlation existed for both horticulture and staple crop state farm selection, but it was stronger for staple crop state farms since there were likely coordination benefits associated with locating large-scale farms nearby one another.

Chart 2: Spatial Propensity Scores for Staple Crop vs. Horticulture State Farms

The propensity score distributions (Chart 2) illustrate that although most woredas are well-matched, there are certain outlier woredas in the treatment group. For sensitivity analysis, these outlier woredas are thrown out. It is important to note that even when we control for all potential observable characteristics that might have influenced site selection for horticulture and staple crop state farms, the pseudo R^2 value is only .4035 for horticulture and .2545 for staple crops. Therefore, there still is a large amount of random variation in the model, supported by the fact that most treatment units have a suitable counterfactual with similar propensity scores.

There are multiple reasons why state farm site selection was partly arbitrary. National planners faced information asymmetries when deciding on site selection. The first large scale study to understand the challenges the state farm sector faced was conducted in 1986 after most state farms were already constructed (MSFD Main Report 1986). The lack of preliminary research conducted on farm site selection is highlighted by the fact that the state farm sector barely outperformed the peasant holding sector despite the Derg committing vast resources.

Therefore, planners likely did not have access to important information that would have influenced site selection. Additionally, widescale commercial farming had only begun during the

Third Five Year Plan (1968-1973). The first commercial farm census was published in 1973, a year before the revolution, and many farms were not yet operational (Directory of Agriculture

1973). Planners did not have information pertaining to the relative production or productivity levels of nationalized commercial farms. Since commercial farming was a recent phenomenon that was driven by foreign direct investment, government institutions also likely lacked the tacit knowledge needed to determine optimal site locations for commercial farms. This reinforces the possibility that alternative locations for state farms likely existed and therefore a propensity score matching technique would be an effective strategy to mitigate selection bias error.

e. Agroprocessing Production

Little empirical research has been conducted to understand the location choices of agroprocessing firms and whether it is related to the agricultural sector. For starters, it is no longer clear that linkages to the agricultural sector are relevant towards the location choices of manufacturing firms. Agroprocessing firms might also make location choices depending on where they source other intermediate inputs and their final consumer market. Krugman (1990) develops the groundbreaking new economic geography model that explains the geographic concentration of manufacturing based on transportation costs and the interaction of economies of scale. Manufacturing firms locate in regions to minimize their transportation costs of intermediate inputs and final goods. The spatial concentration of manufacturing is an endogenous process that has increasing returns to scale as firms produce intermediate inputs for other manufacturing firms. Firms collectively minimize their transportation costs because they locate nearby where they source their inputs, which are other manufacturing firms, and where

41 they sell their final goods, either to other firms or consumer markets. But this model may not be relevant to the agroprocessing sector since their main input comes from the agricultural sector, not other manufacturing firms, and this input is bulky and perishable. Therefore, to minimize transportation costs, it may be more productive to locate by your input supply.

There may be other advantages associated with locating in state farm counties, especially if state farm investments were coupled with road infrastructure investments. Following a spatial regression discontinuity design, Dell & Olken (2017) found the 18th century Dutch sugar cultivation system had spatial path-dependent effects on current Java manufacturing. Since sugar needed to be refined onsite, this caused agroprocessing firms to locate nearby. The government invested heavily in road networks to bring refined sugar to export markets, promoting additional industrial agglomeration that persisted well after the Javanese sugar industry collapsed during the

Great Depression. Manufacturing agglomeration persisted after the sugar market collapsed because public infrastructure reduced transportation costs for other manufacturing firms. The endogenous growth cycle that was documented in Krugman’s paper had already developed; other firms had located nearby these investments and this spurred further agglomeration outside the sugar processing sector.

Therefore, if state farm investments affected the geospatial distribution of commercial farming production and they were also coupled with road infrastructure investments, then they likely also affected agroprocessing firm agglomeration. In order to empirically assess this relationship, I follow the same Pooled OLS and propensity matching approach as the commercial farm equations and aggregate commercial farm data to the county level. The main dependent variable is the logged total production in Ethiopian Birr and the vector of controls include

42 population density, modern road density, distance to historic manufacturing towns, and distance to the capital.

f. Agroprocessing Productivity

Limited research has been conducted on the underlying mechanisms hampering productivity growth within the Ethiopian agroprocessing sector and whether it is attributed to poor linkage formulation with the domestic agricultural sector. Answering how modern linkages affect agroprocessing productivity is complicated. It is challenging to pinpoint the effect agricultural production has on agroprocessing productivity due to simultaneity bias; production growth in the agriculture sector leads to productivity growth in the agroprocessing sector, which leads to increased demand for inputs from the agricultural sector. The underlying mechanisms are also difficult to study because they require extensive data collection on where firms source inputs.

With the advent of global value chains, it is also not clear whether firms make location decisions based on domestic input supplies. Henderson et al. (2016) builds upon Krugman’s model and found late industrialized countries have more spatial inequality than early industrializers since they are more likely to locate in regions that reduce international transportation costs. When early industrializers underwent structural transformation, international transport costs were high and little international trade was conducted. Therefore, to reduce transportation costs, firms located within agricultural heartland regions nearby their input supply. Since manufacturing agglomeration had already begun, falling international transportation costs did not affect the location choices of firms for early industrializers. On the other hand, firms from late industrializers located near coastal regions to reduce the cost of

43 imported intermediate inputs and gain access to export markets since manufacturing agglomeration had not yet begun before international transportation costs fell.

Firms may no longer enjoy the same productivity benefits associated with locating in domestic agricultural heartlands if they source their inputs abroad. The first question we must ask is whether backward linkages to the domestic agricultural sector even influence firm productivity and more broadly, structural transformation. Domestic value chains may still directly improve firm productivity if foreign raw inputs are relatively more expensive than domestic inputs due to high international transportation costs. This is likely true for Ethiopia since it is a land-locked country where 95% of import-export trade flows through Djibouti

(Cheru et al. 2019, 117). As well, supply chains are often interrupted due to inefficient customs systems (Oqubay 2019, 614). Furthermore, poor town integration into foreign markets may stymie structural transformation in second cities, as firms often resort to locating in the capital, where land and labor is relatively more expensive, so they can gain access to imported intermediate inputs.

There may also be indirect benefits associated with creating domestic value chains. The

GTP-II outlines various roles second cities can play, “strengthening trade, building access to regional and international markets, linking emerging sectors, forming clusters that lend themselves to the advantages of agglomeration” (National Planning Commission 2016). Second cities are also more likely to be in the agricultural heartland of Ethiopia, serve regional agricultural markets, and provide employment opportunities for farmers outside the peasant holding sector (Dorosh et. al 2018). In order to investigate this relationship, firm level manufacturing data is used to create a simple Cobb-Douglas production function and vary where firms source their inputs in equation (12).

Y K I 12 ln( ) = β + β ln ( ) + B M + B L + B ln ( 0 ) + X + ε W fwt 0 1 W fwt 2 fwt 3 fwt 4 W fwt wt pwt

The dependent variable for equation (12) is logged labor productivity for firm f in woreda w at time t, measured by output per worker in Ethiopian birr. The main outcome of interest is the

I logged input per worker, ln( 0 ), where I takes the form of either foreign or domestic inputs. We W 0 expect the coefficients for both the domestic and foreign variables to be positive since additional inputs per worker should increase labor productivity. But, if we expect sourcing inputs domestically has a stronger productivity effect, then we expect the coefficient on the domestic inputs per worker variable to be larger than the foreign inputs per worker coefficient. Mfwt and

Lfwt are firm size dummy variables that equal one if a firm employees 51-100 workers or more than 100 workers, respectively. Larger firms are expected to be more productive than smaller firms, so it is important that we control for firm size. It is also important to control for different

K levels of capital stock, which is captured by ln ( ) , the logged capital per worker. X is a W fwt wt vector of present woreda characteristics that may also influence firm productivity such as population density and road density. In order to ensure all continuous logged variables are normally distributed, I throw out all logged observations for the dependent and independent variables that are less than 0. Firm level observations that have zero values for relevant variables is likely misreported and should be thrown out of the analysis.

I K 13 ln( 0) = β + β ln( ) + B M + B L + B ln( C) + X + ε W fwt 0 1 W fwt 2 fwt 3 fwt 4 wt wt pwt

I K 14 ln( 0 ) = β + β ln ( ) + B M + B L + B ln (C′) + X + ε W fwt 0 1 W fwt 2 fwt 3 fwt 4 wt wt pwt

Finally, equation (13) analyzes whether county level commercial farm production influences agroprocessing firms’ access to local inputs and whether this input sourcing advantage

45 translates into productivity gains. The dependent variable of equation (13) is now logged local inputs per worker and the main variables of interest is Cwt, which is the woreda level commercial farm production at time t. If woreda level commercial farm production affects firms’ access to local inputs, we expect Cwt to have a positive sign. Following the rationale outlined above, all logged observations for the dependent and independent variables that are less than zero are thrown out.

But the commercial farm production variable is likely endogenous due to simultaneity bias; not only does commercial farm production influences agroprocessing firm productivity, but agroprocessing firm input demand affects commercial farm production. To weed out this simultaneity bias, a 2SLS approach is used by instrumenting the state farm dummy variables on present commercial farm production. The first stage of the Hausman-Wu (1978) test finds that both the staple crop state farm dummies and horticulture state farms meet the strong and relevancy criterion (See Appendix Table 11). These variables make strong instruments because they only affect the dependent variable through the endogenous variable.

The dummy variable has variation since many counties that had commercial farmland in

1973 did not obtain a state farm investment, and therefore, alternative suitable areas existed that were not treated by state farm selection. This model exploits this variation within the dummy variable to compare counties that have commercial farmland because of state farm investments, to counties that do not, or have much less now, because they were not treated with a state farm investment. Since both instruments are historic variables and only affect the local input supply for agroprocessing firms through its influence on the location of present commercial farm production, these variables should not be correlated with the error term. This variation is used to

46 identify how county level commercial farm production levels influences agroprocessing firms’ access to inputs and labor productivity.

IV. Data Sources Since the paper examines the location choices of commercial farms, their productivity and influence on agroprocessing firms, it was imperative to find plot and firm level data. My analytical approach also requires linking plot level commercial farm and firm level manufacturing data to woreda level geographic, infrastructure, and historic characteristics.

Therefore, it was necessary to georeference all manufacturing firms and commercial farms at the woreda level. Current woreda boundary information comes from the 2013 Open Africa Shape

Files repository and includes 684 Ethiopian woredas located within ten regions and three administrative states (Harar, Dire Dawa, and Addis Ababa). Although woreda boundaries did not change from 2011-2015, historic woreda boundaries from the Imperial and Derg regime have changed, making it challenging to utilize historic data. Multifaceted geospatial and name linking strategies were employed to link relevant historic data to contemporary county boundaries, which are detailed in the “Historic Data” section of the Data Sources description.

a. Contemporary Data

Contemporary commercial farm area, production, productivity, and input data comes from the Central Statistical Agency’s (CSA) Large and Medium Scale Commercial Farms

Sample Survey for 2011, 2013, 2014, and 2015. The CSA did not conduct a commercial farm survey in 2012 nor have they conducted another survey since 2015. The CSA defines a commercial farm as a legally established farm that is profit oriented and uses modern inputs, irrigation schemes, fertilizers, and machinery to attain high plot productivity. Surveyors employed a stratified sampling technique at the region level and planned to survey 2851 farms in

2011, but only managed to cover 2289 farms. In 2013, 2014, and 2015 surveyors planned to sample 3148, 3179, and 3255 farms but only successfully covered 2927, 2987, and 3041 farms

48 respectively. Although the survey stratifies at the region level, it is also representative at the woreda level since the sample covers approximately a quarter of all commercial farms and efforts were made to ensure the stratified regional samples were geographically representative at the sub-regional level. Although surveyors attempted to cover inputs used, only the 2015 survey has systematic data on irrigation and improved seeds at the plot level. I managed to geographically locate 97.8% of commercial farm plots within the sample at the woreda level.

Some plots were missing woreda information and therefore could not be used in the analysis.

Current manufacturing data comes from the Central Statistics Agency Large and Medium

Manufacturing and Electricity Industries Survey (LLMIS) for 2012, 2013, 2014, and 2015. The

CSA covers all manufacturing firms that employ more than ten workers and use power driven machinery. I managed to link 98.3% of all firms covered during this time period to their respective woredas. It is important to note that I aggregated manufacturing firms located in town woredas to their respective rural woredas to ensure conformability with the commercial farm data. Ethiopian towns often have their own woredas but the geographic areas of town woredas are too small to capture surrounding rural areas that might have commercial farming production.

In order to capture potential geospatial relationships between manufacturing and commercial farm production, I used ARCGIS to aggregate each Ethiopian town covered in the CSA manufacturing census to the specific rural woreda that encompasses the town. I did not employ this strategy for major towns like Addis Ababa, Harar, and Dire Dawa. These cities have woreda areas similar to rural woredas. I also categorized firms at the sector level. Food processing firms produced final goods where the main input is a staple crop, horticulture crop, sugar, or water.

Agroprocessing firms include all food, beverage, textile and leather processing firms.

Geographic data used in the analysis comes from a variety of sources and was aggregated to the woreda level using ARCGIS. Soil data comes from the FAO Digital Soil Map of the

World. For eleven different soil types (arcisols, arenosols, xerosols, vertsols, fluvisols, ferralsols, lithosols, planosols, nitosols, regosols, and cambisols) I aggregated the total soil area for each soil type to woreda level and then divided the total area of the soil type. This metric captures the intensity level of each soil type at the woreda level. I also experimented with using the percentage of area each soil group covers in every woreda, but I found the first approach was better suited for my analysis because certain soil types, which are rarely found in Ethiopia, are better for select crops, which is effectively captured by using the intensity approach. The intensity soil variables are not normally distributed and have large left tails since many counties have zero values and only a few counties have a disproportionate share of the total soil area.

Therefore, I transformed the intensity variables into dummy variables. The dummy variables equal one if the intensity level equals or exceeds the 90th percentile. Select soil intensity dummy variables (ferralsols, lithosols, arcisols, and planosols) equal one if the woreda’s intensity level is equal to or larger than the mean, where the mean is between the 90th-95th percentiles. I used the mean instead of the 90th percentile for these select soil types because their distributions are extremely skewed. If I had instead used the 90th percentile, select woredas would have been captured by the dummy variable that have marginal soil intensity levels.

Elevation data comes from the United Nations Office for Coordination of Humanitarian

Affairs (OCHA) and I aggregated the elevation data into three groups using ARCGIS: woreda land area located 0-1000 meters above sea level, 1000-2000 meters, and above 2000 meters. For each woreda, I took the elevation variables’ area and divided it by the woreda’s total area. These metrics capture the proportion of woreda level area located between 0-1000 meters, 1000-2000

50 meters, and above 2000 meters. I then followed a similar strategy to the soil data and transformed the continuous variables into dummy variables that equal one if a woreda’s elevation proportion is equal to or larger than the 75th percentile for each respective elevation group.

River data comes from the United States Geographic Survey World HydroSHEDS

Initiative. River intensity is computed by aggregating total river length in kilometers at the woreda level and then dividing by each woreda’s area in squared kilometers. Contemporary road data comes from 2015 Ethiopia Open Street Map and includes all road types found in the database. Following the same approach as the river intensity variable, road intensity is found by summing the total road length within each woreda and dividing by the woreda’s area. I could not find similar data for 2011-2014 so the contemporary road intensity variable does not change across the duration of the panel. Yet, this should not be problematic because over 75% of the construction on the Road Sector Development Program was already completed by 2010

(Shiferaw et al. 2015, Adamopoulous 2018). 2011-2015 woreda level population density data comes from World Pop, which combines satellite, survey, census, and GIS datasets within a machine learning framework to calculate annual population densities accurate at the 100m by

100m grid cell. These grid cells were aggregated and merged at the woreda level.

b. Historic Data

Historic datasets from 1973-1980 are used to construct the propensity score matching technique model detailed in the Methodology section. For my propensity model, I found woreda level manufacturing and commercial farm data right before state farm investments were implemented in 1979/1980. Town level electricity capacity and production data comes from the

1978 Central Statistics Office Statistical Abstract. This dataset was used to find the straight line

51 distance between woredas and the closest town that had the capacity for manufacturing, right before national planners decided on the site location of state farms. Straight line distances were computed by taking the centroid location of a woreda and the centroid location of the closest town with electricity capacity in 1978 for each respective woreda. Road, mineral deposit, cattle center, and secondary school location data during this time period comes from maps found in the

1981 National Atlas of Ethiopia. This map repository contains a variety of maps dated between

1975-1981 that are relevant for my analysis. Road data comes from the “1979 Ethiopia

Electricity Supply System Map”, an infrastructure map that also contains all-weather and dry- weather road locations. Mineral deposit point locations come from the “1978 Ministry of Mines,

Energy and Water Resources Map” and I grouped mining deposits into four different groups: coal and iron, non-ferrous metallic minerals, and nonmetallic minerals. Cattle center location data was extracted from the “Distribution of Cattle and Veterinary Centers Map”. The secondary school point locations come from the “Ministry of Education 1978 Map”.

In order to utilize the location data, I used ARCGIS to extract the data from maps and aggregate it to the woreda level. Because these maps were generated over 40 years ago and do not include information pertaining to their coordinate projections, I resorted to manually georeferencing each map in ARCGIS. Georeferencing attempts to align a historic administrative boundaries map with another map displaying the current administrative boundaries by taking control points to each map and “pin-pointing” the same feature on both the historic and contemporary maps. Map 2 illustrates the georeferencing process used to align historic maps’ administrative boundaries with contemporary administrative boundaries. The red crosshairs are the point location of a distinguished feature on the historic map, like a border point, and the green crosshair is the same feature located on the contemporary map. The historic map is

52 superimposed onto the contemporary map and is slightly translucent to illustrate the outlines of the contemporary map.

Map 2: Georeferencing Process for Secondary School Locations in 1978

If the historic and contemporary boundaries perfectly aligned with one another, then the red and green crosshairs would perfectly overlap, like crosshairs in the bottom left hand corner of

Map 2. But there are marginal discrepancies. These discrepancies are theoretically problematic because, if large enough, school point locations could be misidentified. When we aggregate the georeferenced data to the woreda level, the centroid of each secondary school is mapped to the specific woreda that encompasses the centroid. Yet, if the control point error is large enough, calculated by taking the distance between the red and green crosshairs, the secondary school

53 might be matched to the wrong woreda. One potential solution to mitigate this error is to follow a

Bayesian approach and take the control point average error and variance, assume a distribution, generate thousands of different aggregate datasets that are shifted by the generated error value from the distribution and run the model thousands of times with the different datasets and average the coefficients across the models. Yet, this is technically challenging in ARCGIS and since the errors are marginal relative to the average size of a woreda, I will assume this potential source of error is not problematic.

Following the same georeferencing strategy explained above, I also constructed woreda level agricultural production variables using map data from the 1982 FAO- UNDP Land Use

Planning Project. These maps illustrate major production areas for barley, coffee, cotton, maize, teff, and wheat. In order to construct intensity variables, I aggregated major production area to the woreda level and divided by the total major production area for each crop type. Major production regions were identified by the FAO as areas where peasant holding agriculture intensively farmed the crop of interest. Following the same approach as the soil type dummy variables outlined in the Contemporary Data section, I then converted these continuous crop type intensity variables into dummy variables since they were not normally distributed. Barley, coffee, maize, teff and wheat dummy variables were constructed using the 90th percentile while the cotton dummy variable was constructed using the 95th percentile. For instance, if a county had a barley intensity value larger than the 90th percentile, then the county’s barley intensity dummy variable value was a “1” while other counties that did not meet this criterion have “0” values.

Commercial farm data from the Imperial regime comes from the 1973 Directory of

Agriculture by the Ethiopian Chamber of Commerce. This census data includes the commercial

54 farm name, location, and crop type grown for 1450 commercial farms. Since this was the first and only edition of the directory, some establishments were likely omitted. Fortunately, most commercial farms in the directory have locations that are the same or very similar to present woreda names. In order to match commercial farms to present woredas, I use a name linking strategy and match 1973 commercial farm locations to present woredas based on name similarity. I also graded each match using a 1-5 scale. Using this name matching strategy, I managed to link 98.6% of commercial farms to present woredas. The criteria for a “5” quality match is an identical name match, or a name match, off by one character. The criteria for a “4” quality match is that the name match can only be off by 2-3 characters. This is quite a common occurrence since woreda names are translated from Amharic to English, which leads to multiple

English translations. 91.5% of matches have a match quality score of 4 or higher. A “3” quality match has a similar woreda name to a present woreda but has more than three different characters. A “2” match quality does not use the name matching technique, but rather uses

Google Maps to geolocate the commercial farm to a present woreda by finding a town or geographic feature that has a similar name as the historic commercial farm location. Finally, a

“1” match quality uses Google Maps to geolocate the commercial farm, but the names of the town or geographic feature are only remotely similar. In total, there was only one commercial farm out of 1450 that had a “1” quality match. Although this linking strategy cannot explicitly account for changing woreda boundaries, this data source is still relevant for my analysis. If historic woreda names are similar to current woreda names, then they likely cover the same area.

Furthermore, this approach allows me to use an important data source that is specifically relevant towards the site selection criteria national planners used when deciding where to locate state farms.

The main variable of interest, a dummy variable denoting whether a woreda had a state farm during the Derg regime, was challenging to construct because the Ministry of State Farms

Development (MSFD) did not keep reliable documentation on the location and operation of state farms. Since farms were organized and operated by regional subsidiaries (Northern, Western

Southern, and Awash) there are very few data sources that cover all state farms. Furthermore, each map or production dataset is only a snapshot in time and does not include state farms that were not yet constructed or state farms that were no longer operational. Therefore, I had to rely on a multitude of national maps, regional maps and production data between 1980-1989 to cross- validate the location of historic state farms with present woredas.

The main data sources used to geolocate state farm counties are the 1981 National Atlas of Ethiopia, 1985-1987 Regional Atlases of Ethiopia (Central, Eastern, Southern, Western, and

Northeastern), and the European Commission to Ethiopia Local Food Security Unit. The 1981

National Atlas of Ethiopia, where the underlying GIS data is provided by the Ministry of State

Farms Development in 1981, gives the point location for 41 state farms. Three state farms were operated by the Northern Development Corporation, six by the Western Development

Corporation, fifteen by the Southern Development Corporation, eleven by Awash Corporation, and five operated by the Horticulture Development Corporation. The 1985-1987 Regional

Atlases of Ethiopia provide the point location of 51 state farms and includes aerial land use maps that outline state farmland area. Eleven state farms were in the Southern region, fourteen in the

Western region, twenty in the Central region with most located in the southern part, three in the

Eastern region, and three state farms in the Northeastern region. Finally, the European

Commission to Ethiopian Food Security published a report in 2000 that includes the present woredas for 26 staple crop state farms, with only half of the farms operational at the time.

I relied on all three GIS sources to properly match state farm investments during the Derg regime to present woredas. I first began by georeferencing the National Atlas of Ethiopia point location map, Regional Atlases of Ethiopia point location maps, and Regional Atlases of

Ethiopia land use maps in ARCGIS using the same strategy outlined for Map 2. Then, I overlaid all three maps with the present woreda boundary map in ARCGIS to assess which present woredas had state farms. By combining GIS information from all three sources, I was able to construct a composite location and gain an accurate picture on the size and location of state farms. The land use maps were especially effective for geolocating staple crop state farms because these farms spanned thousands of hectares and were often located in multiple woredas, which is not captured by point location maps. It is also important to note that I cross-validated my GIS matching strategy with the 26 staple crop state farms found in the European Commission report and found this to be an effective strategy.

I graded each state farm match on a 1-5 scale. State farm matches with a “5” quality match were either found in the European Commission report, found on Google Maps with a present commercial farm having the exact same name as a historic state farm, or the GIS state farm point locations perfectly aligned with the land use map. I define a perfect GIS match as the land use polygon encompassing at least one state farm location point. A “4” quality match had state farm point locations and land use maps that were marginally misaligned, but the GIS data could still be used to identify the woreda. This misalignment is likely attributed to georeferencing error. “3” quality matches have GIS point locations from the National and

Regional Atlases that align closely with one another. A “2” quality match follows the same strategy as “3” matches but the GIS points are farther removed from one another. Finally, a “1” quality match are those that only have one GIS point location. 65.2% of state farm matches have

“5” scores, 9.1% have “4” scores, 12.1% have “3” scores, 10.6% of state farms have “2” scores, and 3% of state farms have “1” scores.

I also used state farm financial and production data to cross-validate the existence of state farms. In a few instances research facilities or cooperative farms were erroneously included in the state farm GIS maps. For instance, out of the 71 potential state farms found across all sources, five were thrown out of the analysis because they were either research facilities or cooperatives. Financial and production data gives added confidence to the matching strategy by eliminating farms that either never became operational or farms operated by different agricultural sectors. Financial performance data comes from the Ministry of State Farms Development

Review and Appraisal of the Financial Performance of State Farms from 1981-1985. This source includes annual revenue and cost data for 44 state farms. Staple crop production data comes from the Southern Agricultural Development Corporation 1986 General Report found in Gabriel

(1990) and includes area, production, cost, and productivity estimates for 19 state farms between

1980-1985. Horticulture state farm production data comes from the 1986/87 and 1988/89

Horticulture Development Corporation Annual Reports, which includes production data for 30 horticulture state farms. I discarded horticulture state farms that did not have at least 500 quintals of produce since they were not relevant to my analysis. These small farms were organized and operated by non-agricultural government institutions to serve ulterior motives, like the Zwai

Prison Horticulture farm operated by the Ministry of Interior.

I created an additional state farm dummy variable that only includes state farms that are cross-validated with area, production or financial data. The state farm production dummy variable equals one if the state farm has been properly geolocated and has at least one year of financial, production, or aerial land use data. For my sensitivity analysis, I run all models using

58 both the original dummy variable and the cross-validated production dummy variable, but there is no meaningful difference. In total, I was able geolocate 65 out of 66 state farms in 65 woredas for my analysis. It is important to note the number of state farms matching the number of woredas is coincidental since certain woredas had multiple state farms while other state farms spanned multiple woredas. Since the state farm production dummy variable is more restrictive, this dummy variable only includes 61 woredas. Finally, I also categorized state farm woredas into staple crop and horticulture subgroups. All farms managed by the Horticulture State Farm

Development or state farms that grew coffee, tea, or cotton were categorized as horticulture state farms while grain, sesame, and cotton state farms were categorized as staple crop state farms.

V. Data Analysis a. Descriptive Statistics Aggregate commercial farm area and production results come from the 2011-2015 CSA

Commercial Farm Reports. Productivity metrics were computed by dividing total production

(qt.) by total area (ha.) and productivity growth estimates take the averages of annual

productivity growth results. Because the CSA does not have commercial farm data for 2012, I

compute 2011-2013 production and productivity growth instead and then divide by two. From

2010-2015, commercial farm output grew for all major crop types, but this growth was mainly

attributed to expanded crop area (Table 4). GTP-I investments within pre-existing cotton and

sugar farms led to a 38.4% and 37.1% annual increase in output, respectively. Staple crops saw

modest annual production growth of 9% as private commercial farms expanded cultivation on

previously unused land. Coffee annual production growth was 6.9% while horticulture annual

production increased by 20.8%.

Table 4, Panel A: 2010-2015 Commercial Farm Area by Crop type (ha.)

Year Staple Horticulture Coffee Sugar Cotton Total

2010 408,175 10,066 56,882 12,166 34,911 535,630

2011 452,244 12,575 75,048 21,100 40,367 616,462

2013 569,826 12,416 113,543 49,582 165,621 932,313

2014 602,314 13,475 112,316 49,639 167,697 972,867

2015 612,081 14,492 108,007 51,089 172,182 983,537

Avg. 528,928 12,605 93,159 45,894 116,156 808,162

Source: CSA Commercial Farm Reports 2011-2015, area is measured in hectares

Table 4, Panel B: 2010-2015 Commercial Farm Production by Crop type (qt.)

Year Staple Horticulture Coffee Sugar Cotton Total

2010 8,393,447 1,203,787 649,962 15,057,859 771,909 27,413,411

2011 9,327,412 2,109,353 906,962 30,459,650 824,702 46,410,553

2013 9,956,774 1,582,407 897,388 59,235,428 2,892799 75,620,987

2014 11,919,953 1,811,199 825,172 59,812,931 3,320,879 79,641,442

2015 12,151,039 1,920,668 799,715 61,778,945 3,368,755 81,934,805

Avg. 10,349,725 2,156,853 1,019,780 56,586,203 2,794,761 62,204,240

Source: CSA Commercial Farm Reports 2011-2015, production is measured in quintals

Table 4, Panel C: 2010-2015 Commercial Farm Productivity by Crop type (qt./ha.)

Year Staple Horticulture Coffee Sugar Cotton Total

2010 20.56 119.59 11.42 1237.70 22.11 51.18

2011 20.62 167.74 12.08 1443.58 20.43 75.28

2013 17.47 127.50 7.90 1194.70 17.47 81.11

2014 19.79 134.41 7.35 1204.96 19.80 81.86

2015 19.85 132.53 7.40 1209.24 19.57 83.31

Avg. 19.66 136.35 9.23 1258.04 19.88 76.97

Source: CSA Commercial Farm Reports 2011-2015, productivity is measured in quintals divided by hectares

Productivity decreased for cotton and coffee crops with annual productivity growth at

-0.6% and -4.5%, respectively. Staple and sugar crops saw marginal productivity growth with

annualized rates of 1.56% and 2.3%. Horticulture was the only crop to have sustained

productivity growth at 8.07%. There is also large crop variation in absolute productivity levels,

which is likely attributed to disparities in irrigation and improved seed use. In 2015, 4.3%, 54%,

98.3%, and 17.8% of commercial farmland area was irrigated for staple, horticulture, sugar, and coffee crops, whereas 64.2%, 69.8%, 98.3%, and 1.3% of commercial farmland area used improved seeds, respectively.

Chart 2: 2011-2015 Logged Plot-Area Distributions (CSA)

Crop plot sizes may also affect productivity variation (Chart 2). The average sugar plot size was 488.5 hectares while the median plot size was 10.49 hectares. Cotton plots have an average size of 200 hectares and a median size of 45.2 hectares. The average and median coffee plot size was 224 and 54.97 hectares respectively, while staple crop plot mean and median sizes were 164 and 12.5 hectares. The production structure of horticulture commercial farms is much

62 smaller in scope than the other crop types. The average horticulture plot size was only 129 hectares with a median of 1.695 hectares. There are likely productivity advantages towards larger farms, but initial investment costs for inputs and irrigation may be cheaper for smaller plot sizes.

Finally, since the plot size medians are all much smaller than the means, large-scale plantation agricultural production structures exist for all crop types.

Table 5: 2012-2015 Imported Intermediate Input Intensity & Value Added by Sector

Year Food & Beverage-Processing Agroprocessing Non-Agro Import Int. (%)| Val. Added (‘000 Birr) Import Int. (%)| Val. Added (‘000 Birr) Import Int. (%)| Val. Added (‘000 Birr)

2012 45.3% | 5,570,603 56.4% | 6,723,592 73.7% | 10,437,415

2013 38.7% | 8,241,229 46.5% | 10,864,589 68.5% | 14,315,134

2014 37.5% | 12,166,128 38.3% | 14,690,458 73.3% | 17,402,850

2015 31.7% | 21,257,826 35.7% | 24,296,726 68.3 % | 22,137,271

Source: CSA LLMIS Report

Imported input intensity and value added data comes from the 2012-2017 CSA Large-

Medium Size Manufacturing Reports. From 2012-2015, food and agroprocessing value added increased substantially. The average annual growth rates for these two sectors were 56.8% and

54.1%, which also coincided by a decrease in import intensity. Sectoral imported input intensity is calculated by taking the total amount of imported inputs by sector, and then divided by the total amount of inputs used by each sector. Agroprocessing value added growth was almost entirely attributed to the food and beverage sector, while textile and leather industries only comprised 12.5% of the agroprocessing value added in 2015. Non-agroprocessing firms’ value added annual growth was 28.6%, while their import intensity remained stable. By 2016, malt liquor firms, which primarily use wheat and sugar as inputs, comprised 37.3% of the food and beverage processing value added. Sugar mills comprised 8.6% of the value addition while grain,

63 bakery, and pasta firms produced another 24%. Water bottling firms’ value addition comprised

18.9% and finally, fruits and vegetable processing companies comprised 7.5% of the food and beverage value addition. Yet, this explosive growth did not persist; between 2016 and 2017, annual value added growth decreased by 9.4% within the food and beverage sector. This is directly attributed to an inadequate supply of raw material as 54.2% of food and beverage firms cited a shortage of raw material as the largest operational problem they faced.

Table 6, Panel A: Plot level Commercial Farm Descriptive Statistics

Variable Obs Mean Std.Dev. Min Max Area (ha). 26748 185.344 4185.114 0 333000 Production (qt.) 24207 12178.5 508000 0 7.28e+07 Improved Seed 7389 .247 .431 0 1 Irrigation 7343 .181 .385 0 1 Horticulture State 26794 .059 .236 0 1 Farm County Staple Crop State 26794 .309 .462 0 1 Farm County Plot Size (< 5ha.) 28587 .277 .448 0 1 Plot Size (5 ha. -50 28587 .426 .494 0 1 ha.) Plot Size (50 ha. - 28587 .104 .305 0 1 100 ha.) Plot Size (>100 ha.) 28587 .193 .395 0 1

When we examine the descriptive statistics of relevant plot level commercial farm variables we find that 30.9% of all commercial farm plots are in staple crop state farm counties while 5.9% are in horticulture state farm counties. Interestingly, 65% of all commercial farm production from 2011-2015 occurred in state farm counties. Most farms have plot sizes between

5-50 hectares, but there are large left and right tails as 27.7% of plots are less than 5 hectares while 19.3% of plots are larger than 100 hectares. Many plots are missing production data and area, while production standard deviations are large due to misreported data that has been coded

64 to 0. These observations are also discarded from the analysis. There are fewer observations for irrigation and improved seed dummy variables because this data was only available in 2015.

Table 6, Panel B: Firm level Manufacturing Descriptive Statistics

Variable Obs. Mean Std. Dev. Min Max Labor Productivity 9299 472000 1620000 0 1.01e+08 Production (Birr) 9420 4.54e+07 1.89e+08 0 4.88e+09 Capital per Worker 9361 348000 881000 0 4.11e+07 Imported Inputs per 7354 173000 784000 0 2.77e+07 Worker Local Inputs per 8899 169000 631000 0 3.91e+07 Worker Small Firms (<50 9639 .742 .438 0 1 workers) Medium Firms (50- 9639 .102 .303 0 1 100 workers) Large Firms (>100 9639 .156 .363 0 1 workers) Horticulture State 9290 .128 .334 0 1 Farm county Staple State Farm 9290 .057 .231 0 1 County Food Processing 9639 .24 .427 0 1 Bev. Processing 9639 .024 .154 0 1 Textile Processing 9639 .044 .206 0 1 Leather Processing 9639 .049 .216 0 1

Firm level manufacturing descriptive statistics are reported in Table 6, Panel B.

Similarly, some firms are missing production, input, and capital stock data and they are discounted from the analysis. We also see strong variation for these variables and all variables that have zero values are thrown out as well. Food, beverage, textile, and leather processing firms comprise 24%, 2.4%, 4.4% and 4.9% of all firms in the panel, respectively. 15.6% of all firms are in staple crop state farm counties while 12.8% are in horticulture state farm counties.

Firms generally source more inputs domestically than abroad while 74.2% of firms employ less than 50 workers. At the regional level, agroprocessing firms are generally more geographically

65 dispersed than nonagro firms, but a substantial portion of production still takes place in the capital. 26.6% of food & beverage firms are in Addis Ababa, 4.9% in Dire Dawa, 10.9% in

SNNP, 34.5% in Oromia, 12.4% in Amhara, and 8.1% in Tigray. Textile and leather processing firms are more concentrated within the capital city with 77.5% of textile firms and 36% of leather processing firms in Addis Ababa. Interestingly, 42.7% of all imported intermediate inputs used by agroprocessing firms are consumed by firms in Addis Ababa. As well, 17.3% of goods exported by agroprocessing firms are from firms located in the capital city.

Table 6, Panel C: County level Descriptive Statistics

Variable Obs Mean Std. Dev. Min Max Horticulture State 2736 .034 .18 0 1 Farm County Staple Crop State 2736 .061 .24 0 1 Farm County Border State Farm 2736 .304 .46 0 1 County Road Density 2736 .383 .683 .006 8.868 Distance to Capital 2736 285.104 147.737 0 869.242 Population Density 2736 3.025 17.55 .011 336.907 River Intensity 2736 .076 .102 0 1.222 0-1000m Elevation 2736 .25 .433 0 1 1000-2000m 2736 .25 .433 0 1 Elevation Above 2000m 2736 .393 .489 0 1 Elevation Arcisols Soil Int. 2736 .054 .226 0 1 Arenosols Soil Int. 2736 .101 .301 0 1 Xerosols Soil Int. 2736 .107 .309 0 1 Vertsols Soil Int. 2736 .102 .303 0 1 Fluvisols Soil Int. 2736 .066 .248 0 1 Ferralsols Soil Int. 2736 .051 .22 0 1 Lithosols Soil Int. 2736 .054 .226 0 1 Planosols Soil Int. 2736 .051 .22 0 1 Nitosols Soil Int. 2736 .108 .311 0 1 Regosols Soil Int. 2736 .101 .301 0 1

Finally, Table 6, Panel C, reports county level demographic and geographic variables between 2012-2015. Only 6.1% of all counties were staple crop state farm counties while 3.4% were horticulture state farm counties. 30.4% of all counties bordered at least one state farm county. We find wide variation among county level road density and population density variables. The standard deviation is twice the size of the mean and over five times larger, respectively. I also include all geographic variables that are used to control for geographic fixed effects. When we analyze the effect state farms have on the geospatial distribution of economic activity, firm level and plot level panels are aggregated up to the county level and then linked with county level characteristic data. For regressions analyzing productivity outcomes, I link county level characteristic data at the firm level and plot level for each year.

67 b. Woreda level Commercial Farm Results

In this section, the regression results are presented from equations (1) and (10) that explores the effect state farm investments during the Derg regime have on the geospatial distribution of commercial farmland and production. Historic state farm investments clearly altered the location of both commercial farmland and production at the county level. But the underlying farm type matters; horticulture state farm counties show stronger path-dependency due to their location criteria and current urbanization trends. Furthermore, recent government incentives have been successful at attracting staple crop commercial farm investment located in low-lying regions. Table 7 uses county commercial farm area (ha.) as the dependent variable and

Table 8 uses county commercial farm production as the dependent variable. All regressions have standard errors clustered at the county level and suppress the output for region, year, and geographic fixed effects.

The variable of interest, “State Farm County”, varies from 1.236-3.359 across area models and 1.372-4.684 across production models. The coefficients’ variation is mainly attributed to selection bias. OLS estimates do not control for Derg planners’ selection criteria, whereas propensity estimates restrict the sample space to only include suitable counterfactual counties. Suitable control counties are more likely to have commercial farm production in comparison to nonrestricted control group counties. Therefore, when we exclude poorly matched counties from the analysis, the coefficient on the “State Farm County” dummy variable drops from 3.359 to 2.005 for the area model. The coefficient drops further from 2.005 to 1.236 when we restrict our analysis to include only counties that have commercial farm production, Y>0. We find a similar effect on the state farm county coefficient for the production model.

Table 7: Commercial Farm Woreda Area (1) (2) (3) (4) (5) (6) OLS OLS OLS RE RE RE Propensity Propensity Propensity Propensity VARIABLES Distance<.2 Area>0

State Farm County 3.359*** 3.809*** 2.004*** 2.005*** 2.139*** 1.236*** (0.421) (0.424) (0.478) (0.478) (0.765) (0.338)

Border State Farm 0.937*** County (0.222)

Population Density -0.530*** -0.479*** 0.171 0.182 -1.293 -0.128 (0.167) (0.164) (0.626) (0.626) (0.913) (0.470)

Road Density 0.287* 0.259 0.0508 0.0452 0.838 0.310 (0.162) (0.160) (0.536) (0.535) (0.771) (0.419)

Distance to Capital -0.000729 4.86e-05 0.00213 0.00213 -0.00164 0.00391 (0.000943) (0.000937) (0.00370) (0.00370) (0.00483) (0.00254)

Region FE Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Geo FE Yes Yes Yes Yes Yes Yes

Constant 1.916*** 1.438** -2.041 4.491 2.029 6.034*** (0.698) (0.704) (1.252) (2.731) (4.128) (1.620)

Observations 2,736 2,736 512 512 276 336 R-squared 0.292 0.309 0.412

Note: The dependent variable, Woreda Commercial Farm Area in Hectares, is logged transformed as are the explanatory variables, Road Density and Population Density, to ensure normality. Woreda clustered standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

Table 8: Commercial Farm Woreda Production (1) (2) (3) (4) (5) (6) OLS OLS OLS RE RE RE Propensity Propensity Propensity Propensity VARIABLES Distance<.2 Area>0

State Farm County 4.684*** 5.307*** 2.708*** 2.708*** 2.975*** 1.372*** (0.567) (0.566) (0.675) (0.675) (1.015) (0.406)

Border State Farm 1.298*** County (0.316)

Population Density -0.76*** -0.69*** 0.480 0.475 -2.004* 0.160 (0.236) (0.232) (0.904) (0.903) (1.193) (0.593)

Road Density 0.462* 0.424* -0.104 -0.101 0.854 0.203 (0.236) (0.234) (0.713) (0.712) (0.971) (0.484)

Distance to Capital -0.0023* -0.00126 -9.67e-05 -9.55e-05 -0.00398 0.00112 (0.00135) (0.00135) (0.00512) (0.00512) (0.00648) (0.00344)

Region FE Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Geo FE Yes Yes Yes Yes Yes Yes

Constant 3.116*** 2.453** -1.870 9.144** 3.254 11.83*** (1.025) (1.011) (1.641) (3.943) (5.912) (2.377)

Observations 2,736 2,736 512 512 276 333 R-squared 0.287 0.301 0.388

Note: The dependent variable, Woreda Commercial Farm Production in Quintals, is logged transformed as are the explanatory variables, Road Density and Population Density, to ensure normality. Woreda clustered standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

From 2011-2015, only 36% of counties had any commercial farm production. Since suitable counterfactual counties are also less likely to have commercial farmland in comparison to state farm counties, by removing these counties from the analysis the coefficient drops further.

When controlling for geographic, year, and region fixed effects and restraining our sample space to include counties that currently have commercial farm production, we find that counties that

70 had a state farm during the Derg regime now have at least 123.6% more commercial farm area and 137.2% more commercial farm production in comparison to suitable counterfactual counties.

Across all models, these results are significant at the less than 1% significance level. The sensitivity analysis reinforces our findings that state farm investments influenced where current commercial farming is located, even when we restrict our analysis to compare state farm counties against equally suitable alternative locations.

Road density and population density also play an important role in explaining the county level distribution of commercial farmland. Holding all else equal, a one percent increase in population density leads to a 0.53% decrease in commercial farmland area and a 0.76% decrease in production, whereas a one percent increase in road density (km/km^2) leads to a 0.287% increase in farmland and a 0.462% increase in production. These results align with our expectations because as population density increases, there are more smallholder peasant farms and it becomes harder to expand commercial farm area or form new commercial farm enclaves, while denser road networks improve market access. When analyzing the relevancy of other control variables, we should only consider OLS results because when we restrict our analysis using propensity models, we restrict our sample to only include counties with similar characteristics. The control variables lose statistical significance and in some cases the coefficient signs switch. Yet, this does not mean that these variables are no longer relevant.

The “Border State Farm” county dummy variable is both positive and significant, suggesting, even after controlling for geographic characteristics, being located next to a state farm county increases a county level commercial farm area by 93.7% and production by 129.8%.

Yet, it is unclear whether this dummy variable captures an agglomeration effect or georeferencing error. When georeferencing state farm locations to current state farm counties,

71 not all state farms had land use maps (see Data Sources for detailed explanation). Therefore, state farm point locations may have not captured state farmland area located in adjacent counties.

Because of this problem, this coefficient is difficult to interpret.

Table 9: Commercial Farm Production: Staple vs. Horticulture State Farms (1) (2) (3) (4) (5) (6) OLS OLS OLS OLS RE RE Propensity Propensity Propensity Propensity VARIABLES Area>0 Area>0

Staple State Farm 3.193*** 1.637*** 0.389 County (0.642) (0.611) (0.566)

Horticulture State 2.334*** 2.016** 1.654* Farm County (0.526) (0.859) (0.877)

Population Density -0.67*** -0.357*** -0.142 -0.594 -0.969 -1.891* (0.188) (0.136) (0.701) (0.753) (0.828) (1.116)

Road Density 0.540*** 0.262* -0.246 -0.0383 0.0437 -0.355 (0.189) (0.143) (0.620) (0.644) (0.653) (1.083)

Distance to Capital -.004*** -.0034*** -.000609 -.00804* -.00251 -.0114** (0.00119) (0.000892) (0.00475) (0.00418) (0.00528) (0.00556)

Region FE Yes Yes Yes Yes Yes Yes Year FE Yes Yes Yes Yes Yes Yes Geo FE Yes Yes Yes Yes Yes Yes

Constant 2.639*** 1.724** -2.467 -1.284 1.777 2.740 (0.899) (0.740) (1.619) (1.765) (2.112) (3.944)

Observations 2,736 2,736 424 344 264 203 R-squared 0.220 0.181 0.415 0.348

Note: The dependent variables, Woreda Commercial Farm Staple & Horticulture Production in Quintals, is logged transformed as are the explanatory variables, Road Density and Population Density, to ensure normality. Woreda clustered standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

When we split the variable of interest, state farm county, into horticulture and staple crop state farm county dummy variables, this location effect is stronger for horticulture state farm counties (Table 9). The coefficient on the models with the horticulture state farm dummies range from 1.654-2.334 and they are all statistically significant at the less than 1% significance level.

The coefficients on the staple crop state farm dummies range from 0.389-3.189 and the coefficient on the propensity model, Y >0, is not significant. These different outcomes are related to the location criteria horticulture and staple crop farms use and corresponding urbanization trends. The main criteria national planners used when locating horticulture state farms was road density, access to a rural labor supply, and being in the highlands region so farms could easily access the capital city (Methodology, Table 3). Staple crop state farms were generally located in the unpopulated middle and low-lying reasons where huge tracts of commercial farmland could be grouped together to form large-scale state farms with minimal displacement to peasant holding farmers (Table 3).

Map 3: Populated Density (Ethiopia’s Spatial Structural Transformation)

Since the inception of state farms, these demographic trends have been compounded as middle and low-lying regions remain unpopulated while the highland regions have become densely populated (Map 3). Present staple crop commercial farms can still locate in other low- lying counties where large tracts of land are available. From 2010 to 2015, staple crop commercial farms expanded cultivation area by 203,906 ha. whereas horticulture commercial farm area only expanded by 4,426 ha. (Table 4). Now, 15.4% of staple crop commercial farmland is located more than 400 km from the capital city in comparison to only 6.3% of horticulture commercial farmland. Because we see a weaker path-dependent relationship for staple crop state farm counties, these results indicate recent government FDI incentives have been effective at creating new staple crop commercial farm enclaves in low-lying regions.

74 c. Plot level Commercial Farm Productivity Results

This section reports the plot level productivity results from equations (2) and (11). We also find large differences in plot level productivity outcomes for plots located in staple crop state farm counties in comparison to horticulture state farm counties (Table 10). As well, we find that different commercial farm agglomeration patterns for staple crop and horticulture farms arise from different crop productivities. After controlling for county characteristics, year fixed effects, plot level crop type and plot size, we find that plots located in staple crop state farm counties are 32.7% less productive than commercial farm plots in other counties. This result is significant at the less than one percent significance level. On the other hand, plots located in horticulture state farms are 32.1% more productive than other plots and this result is marginally significant with a p-value of 0.108. When we also control for plot level irrigation and improved seed use, the staple crop dummy loses its statistical significance and the coefficient becomes zero. These results suggest productivity disparities between plots located in staple crop counties is almost entirely attributed to irrigation allocation.

County characteristics like road density, population density and even plot size all have a negligible effect on plot level productivity outcomes. But being located far away from the capital is consequential. Commercial farm plots located more than 400 km away from the capital are

32.7%-41.2% less productive than plots located less than 200 km from the capital, further highlighting how difficult it is to expand commercial farmland in remote areas that are further away from the capital. Interestingly, when we run the propensity model for plot level productivity, the coefficients on the staple crop and horticulture state farm counties increase in absolute value. This suggests the propensity technique is not an effective model when analyzing

75 productivity models at the plot level since we expect the coefficients on the variables of interest to decrease when running the propensity model.

Table 10: Commercial Farm Plot Level Productivity

(1) (2) (3) (4) (5) (6) VARIABLES Staple Hort. Staple Hort. Staple Hort. OLS OLS OLS OLS Propensity Propensity

Staple State Farm County -0.327*** -0.0340 -0.0138 (0.108) (0.0803) (0.101)

Horticulture State Farm 0.321 0.142 0.212 County (0.200) (0.128) (0.144)

Plot Area (5-50 hectares) -0.0243 -0.0394 -0.0459 -0.0461 -0.0309 -0.0256 (0.0479) (0.0495) (0.0519) (0.0520) (0.0569) (0.0570)

Plot Area (50-100 0.0108 -0.00575 -0.0271 -0.0288 0.175 0.177 hectares) (0.0686) (0.0681) (0.0808) (0.0805) (0.111) (0.111)

Plot Area (100< hectares) 0.115 0.0811 0.0604 0.0571 0.190 0.191 (0.0955) (0.0913) (0.0843) (0.0840) (0.143) (0.142)

Road Density 0.0877 0.0508 -0.137* -0.142* 0.218 0.220 (0.0932) (0.0990) (0.0800) (0.0802) (0.178) (0.166)

Population Density (.3-2) 0.121 0.0899 0.289 0.294 -0.281 -0.257 (0.165) (0.155) (0.201) (0.195) (0.278) (0.280)

Population Density (2<) 0.0795 0.0611 0.386 0.382 -0.586 -0.566 (0.199) (0.206) (0.254) (0.246) (0.361) (0.368)

Distance to Capital 0.187* 0.166* -0.0158 -0.00409 -0.0843 -0.0500 (200-400 km) (0.0954) (0.0937) (0.0958) (0.1000) (0.170) (0.148)

Distance to Capital (400+ -0.412*** -0.324** -0.55*** -0.52*** -0.0935 -0.0981 km) (0.150) (0.144) (0.148) (0.147) (0.212) (0.208)

Irrigation 1.661*** 1.666*** 1.538*** 1.574*** (0.140) (0.137) (0.250) (0.252)

0.529*** 0.529*** 0.416*** 0.415*** Improved Seeds (0.0857) (0.0855) (0.133) (0.132)

Crop FE Yes Yes Yes Yes Yes Yes Region FE Yes Yes Yes Yes Yes Yes Year FE Yes Yes 2015 2015 2015 2015

Constant 2.254*** 2.311*** 1.588*** 1.596*** 1.581** 1.449** (0.238) (0.237) (0.238) (0.232) (0.717) (0.702)

Observations 20,434 20,434 5,750 5,750 2,790 2,790 R-squared 0.487 0.484 0.674 0.674 0.798 0.799 Note: The dependent variable, Plot Level Productivity (quintals per hectare), is logged transformed as is the explanatory variable Road Density to ensure normality. Logged area or productivity values that were less than 0 were thrown out. Woreda clustered standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

To understand why these productivity disparities exist today between horticulture and staple crop state farm counties and whether they are related to state farm development, I analyzed the investment decisions of state farm planners during the Derg regime (1974-1990). In

1985, only 0.54% of cultivated land was irrigated and irrigation schemes were exclusively used for cotton and horticulture (MSFD Report Volume IV 1986, 1). Irrigation schemes were initially constructed by private horticulture farms during the Imperial regime (1942-1974) and they used surface irrigation methods like furrow and basin systems (MSFD Report Volume IV 1986, 3).

State farm planners did not invest in new irrigation schemes like sprinkler and drip irrigation systems because they were costly, approximately $7,500-$10,000 per hectare in 1978 dollars, and required foreign currency to import. Ultimately, state farm planners elected to support high- value cash crop exports and only twelve farms that had access to irrigation were cotton and horticulture farms (MSFD Report Volume IV 1986, 4).

Table 11: 1980-1985 Select Staple Crop State Farm Productivity (qt./ha.)

State 1980 1981 1982 1983 1984 1985 Farm

Dixis 693.26 697.37 538.62 578.1 935.77 823.44

Lole 1087.85 1086.79 1115.67 994.84 1337.22 1333.08

Adele 718.63 482.22 844.12 550.84 805.11 1010.99

Garadela 858.22 738.74 823.58 777.06 583.3 747.4

Goffer 898.64 634.86 688.08 591.35 563.91 715.87

Hereroh 970.87 745.28 1068.54 739.33 688.37 815.63

Sheneka 203.74 353.44 524.99 353.91 103.40 337.93

Sirufta 1218.29 1014.26 1053.74 618.9 1074.33 1156.11

source: Gabriel 1990, 83

Staple crop state farms, which relied on rain fed agriculture, had land drainage, soil

erosion, and water supply problems since the land was, “dependent on the vagaries of the

climate” (MSFD Report Volume IV 1986, 4 & Gabriel 1990, 17). Staple crop state farm

productivity was extremely variant, and in many cases, annual productivity changed by over

100% (Table 11). Once the government began privatizing state farms, they effectively shifted the

irrigation investment cost burden onto private commercial farms without implementing any

reform measures to reduce the transaction costs associated with implementing large-scale

irrigation schemes. In 2015, only 3% of cultivated land area was irrigated with most of the

irrigated area intended for sugar, cotton, and horticulture. In summary, large productivity

differences between plots located in horticulture and staple crop state farm counties is driven by

irrigation allocation decisions made by state farm planners during the Derg regime. These

78 disparities exist today because there has been no attempt to support commercial farm investment with irrigation schemes.

To investigate why horticulture agglomeration patterns exhibit stronger path-dependency with horticulture state farms, we include an interaction term within the model (see Appendix

Table 17). This variable interacts the “horticulture plot” dummy variable with the “distance to capital larger than 400km” dummy to find the average horticulture plot productivity in low-lying regions. Even when we control for irrigation and improved seeds, these plots are 118% less productive than plots located closer to the capital city. On the other hand, staple crop commercial farmland in low-lying regions (more than 400 km from the capital city) are 50% more productive. Although certain low-lying regions have the agroclimatic suitability for horticulture production, they have had limited success attracting investment due to productivity concerns.

These productivity deficiencies are likely attributed to additional supply chain investments that are needed for successful horticulture production. Export oriented horticulture production requires packaging, cold storage, phytosanitary inspection and cold air cargo transport (Wiersinga and Jager 2009). These supply chain investments are costly and spatially concentrated within highlands regions, which have easier access to export markets. As well, further expansion in low-lying regions is difficult since it is harder to source a consistent supply of labor in unpopulated areas. Due to productivity concerns in low-lying regions, it seems unlikely for export-oriented horticulture enclaves will form there without government investment in supply chain infrastructure. Furthermore, without an institutional framework in place to lease peasant holding land in highlands regions, horticulture farms have found it difficult to expand into productive highland regions that already have the necessary supply chain infrastructure in place.

79 d. Woreda Level Manufacturing Production Results

This section uses the same analytical approach as the county level commercial farm production regressions, but the dependent variable is county level food and agroprocessing production. We find that state farm investments also altered the county level distribution of food and agroprocessing production (Table 12), but horticulture state farms have stronger spatial linkages. By running the random effects model and restricting the sample to only counties that had production, Y>0, state farm counties have 120% more agroprocessing production and 114% more food-processing production. These results are statistically significant at the less than 1% level across all model selections. Interestingly, road density is important in explaining the current county manufacturing production, but even after controlling for its effects, firms still locate within state farm counties. Production likely locates nearby where agroprocessing firms source their raw material since agroprocessing intermediate inputs are bulky and perishable. Since state farms altered the distribution of commercial farmland, they also affected where production occurs.

When we disaggregate the dummy variable into horticulture state farm counties and staple crop state farm counties, we find that the effect is much stronger for horticulture state farm counties (see Appendix Table 21). We ran the random effects model using the propensity matches and restricted our selection to counties that only have food-processing production. We find that horticulture state farm counties have 196% more production in comparison to staple crop state farm counties that only have 32.9% more production in comparison to suitable counterfactuals. The horticulture result is significant at the less than one percent level whereas the staple crop result is statistically insignificant. In the next section, we analyze how firm level productivity differentials explain these agglomeration patterns.

Interestingly, there is also strong path dependency associated with where production is located. The coefficient on the “Distance to nearest town in 1978 with electricity” is negative and larger than one for both agroprocessing and food-processing OLS models. This suggests that agroprocessing still locates in towns that have a history of manufacturing before state farm investments were implemented. But, there has been a shift in relative production between towns located in state farm counties versus non-state farm counties.

Table 12: Woreda Manufacturing Production Results

(1) (2) (3) (4) (5) (6) VARIABLES Agroprocessin Food- Agro Food RE Agro RE Food g Processin Propensit Propensit Prod>0 Prod>0 g y y

State Farm 2.946*** 2.659*** 3.887*** 3.452*** 1.201** 1.138*** County (0.854) (0.848) (1.012) (0.991) (0.466) (0.419)

Population -0.357 -0.278 1.103 1.336 -0.572* -0.574* Density (0.242) (0.235) (0.899) (0.894) (0.295) (0.300)

Road Density 1.686*** 1.524*** 0.497 0.226 1.367*** 1.293*** (0.295) (0.286) (0.856) (0.869) (0.335) (0.361)

Distance to -0.00153 -0.00160 -0.00965 -0.0103 -0.001 -0.00207 Capital (0.00189) (0.00175) (0.00645) (0.00627) (0.00197 (0.00188 ) )

Distance to -1.595*** -1.328*** -1.399*** -0.901* -0.35*** -0.202* nearest town in (0.249) (0.247) (0.511) (0.508) (0.123) (0.114) 1978 with electricity

Year FE Yes Yes Yes Yes Yes Yes Region FE Yes Yes Yes Yes Yes Yes

Constant 6.726*** 6.230*** 9.536*** 7.056** 21.34*** 20.51** *

(1.929) (1.825) (2.622) (2.890) (1.220) (1.140)

Observation 2,736 2,736 512 512 440 386 s R-squared 0.233 0.204 0.292 0.249

Note: All Continuous Variables are logged. Agroprocessing firms include all food, textile, beverage, and leather processing firms. Woreda clustered standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1 e. Firm Level Labor Productivity Results

This section reports the manufacturing labor productivity regression results for equations

(12), (13), and (14). Furthermore, we analyze the underlying mechanisms that drove value added growth within the food-processing and agroprocessing sector from 2012-2015 (Table 5) to understand whether growth was attributed towards domestic or international intermediate input channels. Afterwards, we analyze whether commercial farm sector production growth during this time period led to improvements in agroprocessing labor productivity. Finally, we end our analysis discussing why agro-processing firms conglomerate within horticulture state farm counties and not within staple crop state farm counties.

We find that firms with strong backward linkages to the domestic agricultural sector are more productive than firms that import intermediate inputs, but input sourcing effects are heterogenous across sectors (Table 13). The main variables of interest for these models are

“Local Inputs per Worker” and “Imported Inputs per Worker”. We run these models separately because the variables of interest are multicollinear. The effect is strongest for food processing firms; a one percent increase in local inputs per worker leads to a 0.222% increase in labor productivity, whereas a one percent increase in imported inputs per worker only leads to a

0.035% increase in labor productivity. Both results are significant at the less than 1% significance level across all model selections. This effect is weaker for non-agro processing firms

82 as a one percent increase in local inputs per worker leads to a 0.153% increase in labor productivity whereas a one percent increase in imported inputs per worker leads to a 0.12% increase in labor productivity. This makes sense nonagro firms may require capital intensive intermediate inputs that are not produced in Ethiopia.

Table 13: OLS Intermediate Inputs Effect on Firm Labor Productivity (1) (2) (3) (4) (5) (6) (7) (8) VARIABLES Non- Non- Agroprocessing Agroprocessing Bev, Bev, Food Food Agro Agro Textile & Textile & Processing Processing Leather Leather Processing Processing

Capital per Worker 0.712*** 0.713*** 0.649*** 0.824*** 0.665*** 0.639*** 0.639*** 0.884*** (0.0320) (0.0427) (0.0563) (0.0321) (0.0687) (0.0769) (0.0736) (0.0306) Local Inputs per Worker 0.153*** 0.207*** 0.173*** 0.222*** (0.0249) (0.0383) (0.0527) (0.0474) Medium Firms (50-100) 0.146*** 0.0162 0.121** 0.176** 0.0591 0.191 0.167*** 0.267*** (0.0536) (0.0481) (0.0498) (0.0810) (0.102) (0.133) (0.0545) (0.0846) Large Firms (100<) 0.275*** 0.0761* 0.107** 0.0766* 0.0919 0.152* 0.115 0.179* (0.0515) (0.0442) (0.0518) (0.0399) (0.0741) (0.0897) (0.0796) (0.0960) Pop. Density (5-25) 0.0188 -0.0561 -0.0106 0.101 -0.116 0.335 -0.0184 0.115 (0.0526) (0.0477) (0.0826) (0.0941) (0.160) (0.205) (0.0900) (0.0829) Pop. Density (25-100) -0.140** -0.165* -0.0763 0.188 -0.173 0.344 -0.0465 0.329*** (0.0595) (0.0842) (0.110) (0.139) (0.244) (0.271) (0.111) (0.0935) Pop. Density (100+) -0.119 -0.117 -0.132 0.207 -0.0737 0.544* -0.173 0.216 (0.0753) (0.0954) (0.119) (0.126) (0.287) (0.279) (0.136) (0.133) Road Density 0.0471* 0.0256 0.0511 0.0322 -0.123 -0.0795 0.0943** 0.0382 (0.0267) (0.0294) (0.0398) (0.0599) (0.115) (0.114) (0.0414) (0.0582) Imported Inputs per 0.120*** 0.0640*** 0.165*** 0.0345*** Worker (0.0229) (0.0144) (0.0433) (0.0128) Constant 1.997*** 2.473*** 2.204*** 1.715*** 2.324*** 2.794*** 2.153*** 1.208*** (0.203) (0.281) (0.338) (0.290) (0.444) (0.543) (0.455) (0.294)

Observations 4,906 3,034 2,878 1,548 840 595 2,038 953 R-squared 0.779 0.816 0.753 0.765 0.696 0.666 0.764 0.831 Note: All Continuous Variables are logged and logged values less 0 are thrown out. Agroprocessing firms include all food, textile, beverage, and leather processing firms. Woreda clustered standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

Interestingly, the input sourcing effect is weakest for beverage, textile and leather processing firms. With domestic and imported input per worker coefficients equaling 0.173 and

0.165, their effects are essentially the same. Although cotton has been grown on large plantations since the Imperial regime and Ethiopia has one of the largest stocks of cattle in the world, backward linkages to these sectors remain weak. Oqubay (2016) documents the specific drawbacks these sectors face and explains why firms have been unable to capitalize on

Ethiopia’s input comparative advantages. These problems mainly revolve around the supply of quality inputs as tanneries are unable to procure high quality leather and textile firms cannot obtain the necessary quantity of raw material needed to be the main raw material suppliers for export-oriented apparel firms. Sectoral variation highlights that not all backward linkages are created equal. Furthermore, not all manufacturing sectors that would theoretically benefit from taking advantage of a country’s comparative advantage are productive.

Across all models, capital per worker has the largest effect on labor productivity. All else equal, a one percent increase in capital per worker improves labor productivity by 0.639%-

0.884%. This effect is stronger for firms that import inputs and for non-agro manufacturing firms. In line with our expectations, larger firms are more productive than smaller firms and increased road density improves labor productivity. Yet, firms located in large cities like Addis

Ababa, Dire Dawa, and Harar are respectively 21%, 10.7%, and 23.8% less productive than agroprocessing firms located in the Tigray region (see Appendix Table 19). Generally, large urban centers are expected to have higher total factor productivities. But, considering agroprocessing firms are more productive when they source inputs domestically and we have established that agroprocessing firms locate within state farm counties, manufacturing value

added growth during this time period may have been attributed to firms having more inputs per

worker from increased county level commercial farm production.

Table 14: Instrumental Variable (IV) Local Inputs per Worker (1) (2) (3) (4) VARIABLES Agroprocessing Agroprocessing Food Food Processing Processing

Commercial Farm Prod. 0.0341 0.0495 0.0353 0.0432 (0.0255) (0.0522) (0.0248) (0.0610)

Capital per Worker 0.911*** 0.909*** 0.880*** 0.880*** (0.0412) (0.0403) (0.0603) (0.0600)

Medium Firm (50-100) -0.149* -0.149* -0.0772 -0.0812 (0.0779) (0.0768) (0.111) (0.121)

Large Firm (100+ workers) -0.282*** -0.283*** 0.0559 0.0506 (0.0734) (0.0708) (0.122) (0.127)

Road Density 0.0672 0.0669 -0.00105 -0.00239 (0.0585) (0.0643) (0.0817) (0.0868)

Region FE Yes Yes Yes Yes Year FE Yes Yes Yes Yes

Constant 0.169 0.146 0.564 0.542 (0.481) (0.499) (0.746) (0.803)

Observations 2,919 2,919 2,064 2,064 R-squared 0.525 0.521 0.524 0.522 Note: Regressions 1 and 3 use the Horticulture State Farm dummy variable as the instrument, whereas regressions 2 and 4 use the Staple Crop State Farm county as the instrument. All Continuous variables are logged and logged values less than 0 are thrown out, besides commercial farm production. Agroprocessing firms include all food, textile, beverage, and leather processing firms. Woreda clustered standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

Table 15: IV Firm Labor Productivity (1) (2) (3) (4) VARIABLES Agroprocessing Agroprocessing Food- Food- Processing Processing

Commercial Farm Prod. -0.00741 -0.0399 0.00326 -0.0483 (0.0144) (0.0265) (0.00947) (0.0327)

Capital per Worker 0.840*** 0.839*** 0.828*** 0.820*** (0.0303) (0.0311) (0.0465) (0.0481)

Medium Firm (50-100) 0.111** 0.104* 0.185*** 0.201*** (0.0517) (0.0559) (0.0543) (0.0709)

Large Firm (100+ workers) 0.0465 0.0392 0.124 0.151* (0.0469) (0.0484) (0.0823) (0.0812)

Pop. Density (5-25) -0.00795 -0.200 0.0573 -0.236 (0.116) (0.202) (0.110) (0.241)

Pop. Density (25-100) -0.0675 -0.277 0.0385 -0.263 (0.144) (0.261) (0.158) (0.314)

Pop. Density (100+) -0.137 -0.444 -0.122 -0.588 (0.165) (0.344) (0.183) (0.439)

Road Density 0.0719 0.117 0.108** 0.181 (0.0485) (0.109) (0.0544) (0.134)

Constant 2.278*** 2.425*** 2.396*** 2.706*** (0.392) (0.438) (0.608) (0.679)

Observations 3,098 3,098 2,104 2,104 R-squared 0.716 0.704 0.721 0.701 Note: Regressions 1 and 3 use the Horticulture State Farm dummy variable as the instrument, whereas regressions 2 and 4 use the Staple Crop State Farm county as the instrument. All Continuous variables are logged and logged values less than 0 are thrown out, besides commercial farm production. Agroprocessing firms include all food, textile, beverage, and leather processing firms. Woreda clustered standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

Tables 14 and 15 show the results from analyzing these underlying mechanisms. County level commercial farm production has only marginally increased the amount of local inputs firms’ source, but this effect does not translate into improved labor productivity. The models use either the horticulture state farm dummy or staple crop state farm dummy variable as the instrument to control for simultaneity bias and isolate the effect county commercial farm production has on manufacturing labor productivity. Holding all else equal, a one percent increase in county level commercial farm production only leads to a 0.034-0.045% increase in local inputs per worker for agroprocessing firms located within the same county, which is significant at the 10% level. But the effect is smaller for food processing firms, only a 0.035-

0.043% increase in inputs per worker, and the result is not statistically significant. Food- processing firms located in counties with denser road networks can source more inputs per worker, but this result is not statistically significant. Since firm size dummy variables are small in absolute value and statistically insignificant, firm size appears to play no role in input sourcing; small firms are just as likely to obtain inputs as large firms.

Commercial farm production coefficients for the instrumental variable labor productivity models are near zero and statistically insignificant (Table 15). The marginal benefits associated with firms obtaining more inputs per worker does not translate into productivity benefits. This is not surprising considering the linkages between the commercial farm sector and the agroprocessing sector are very weak in the OLS models (see Appendix Table 18). This is illustrated by the fact that the woreda level commercial farm production coefficient is almost zero and insignificant. Even when we relax the restrictions for potential simultaneity bias, we see no relevant productivity effect in OLS models. These results clearly indicate that the commercial farm sector did not play a vital role in spurring agroprocessing firm productivity growth.

Interestingly, when we throw out county level commercial farm production in OLS models and instead include the horticulture and staple crop state farm dummies, we find that food-processing firms located in horticulture state farm counties have 24.2% more inputs per worker while food-processing firms located in staple crop state farm counties only have 8% more inputs (see Appendix Table 18). The horticulture dummy is significant at the 10% level while the staple crop dummy is not significant. These additional sourcing benefits only turn into marginal productivity gains for firms located in horticulture state farms, 2%, and this result is not statistically significant (See Appendix Table 20).

On the other hand, food-processing firms located in staple crop state farm counties are

20% less productive and this result is significant at the 10% level. Although firms gain input advantages when locating in horticulture state farms, these advantages do not systematically improve labor productivity. For staple crop state farm counties, the marginal input sourcing gains do not offset other factors driving down labor productivity. It is not clear what is driving this process since even when we control for geographic fixed effects, distance to the capital, and demographic characteristics, the dummy coefficient is negative and statistically significant. But we do know that input sourcing and labor productivity differences between staple crop state farms and horticulture state farms likely explains why horticulture state farm counties now have significantly more manufacturing production than staple crop state farm counties.

VI. Conclusion

In my paper, I analyze the consequences Ethiopian state farm investments have had on the commercial farm and agroprocessing sector. I find that horticulture state farms still largely shape commercial farm and agroprocessing agglomeration patterns because of historic infrastructure investments that are necessary for horticulture production. Horticulture state farms require investments in irrigation schemes and cold storage supply chains. Current horticulture commercial farms conglomerate in these counties to take advantage of preexisting infrastructure since this infrastructure is relatively scarce. Furthermore, agroprocessing firms have conglomerated within horticulture state farm counties to locate by their input supply.

On the other hand, staple crop state farms have not influenced agglomeration patterns because there are no incentives for commercial farms to locate within these historic state farm counties; Derg planners did not invest in infrastructure schemes for staple crop farms and current commercial farm owners have found these investments to onerous and have instead adopted low cost, low output models. Recent FDI incentives for commercial farming in low-lying regions have successfully formed new commercial farm enclaves for staple crops and have exhibited higher productivity levels than staple crop state farm counties, primarily driven by investments in irrigation schemes. Yet, it is unlikely staple crop commercial farm agglomeration will promote agroprocessing agglomeration within these regions since staple crops have weaker spatial linkages to the agroprocessing sector.

My research also finds that agroprocessing linkages with the domestic agricultural sector boosts productivity growth, but linkages with the commercial farm sector remain weak and have not attributed to the recent growth in the agro-processing sector. This is largely attributed to a

90 policy environment that gives exclusive focus to smallholder agriculture. This trend continues as the agroprocessing park industrialization strategy relies on raw inputs provided by smallholder farmers. With growing concerns that smallholder agriculture will not meet these demands and wheat exports expected to surge 37.5% from 2010-2025, private commercial farming could serve as a complementary mode of production to supply intermediate inputs and close the food import gap (Mendes et al. 2015, Will 2019). Yet, without the government creating a legal framework for existing commercial farms to lease land from smallholders and access to credit to expand irrigation schemes, horticulture farms will not be able to expand within the high-lands region.

Furthermore, if the government does not direct investment within supply chain and irrigation infrastructure in low-lying areas, high value added agriculture will continue to be trapped in the highland regions and be defined by its historic legacy.

Sources

“2013 Ethiopia Woreda Shape File.” Open Africa, 29 Jan. 2016. https://open.africa/dataset/ethiopia-shapefiles/resource/7c04c1b9-1e4b-407a-998f- 61276fe721a3

A Homegrown Economic Reform Agenda: A Pathway to Prosperity. Federal Democratic Republic of Ethiopia: Office of the Prime Minister, 2019, A Homegrown Economic Reform Agenda: A Pathway to Prosperity.

Adamopolous, Tasso. Spatial Integration, Agricultural Productivity, and Development: A Quantitative Analysis of Ethiopia's Road Expansion Program. International Growth Centre, 2018, Spatial Integration, Agricultural Productivity, and Development: A Quantitative Analysis of Ethiopia's Road Expansion Program, www.theigc.org/wp- content/uploads/2018/10/Adamopoulos-2018-Working-paper.pdf.

“AGSS Key Findings of 2015/2016.” Central Statistical Agency Agricultural Sample Surveys. pp. 1–22.

Asfaw, Alemu, and Jose Luis Vivero Pol. Is There a Comparative Advantage to Food-Grain Producing State Farms in Ethiopia? European Commission to Ethiopia, 2000, Is There a Comparative Advantage to Food-Grain Producing State Farms in Ethiopia?

Bachewe, et al. “Non-Farm Income and Labour Markets in Rural Ethiopia.” Ethiopian Strategy Support Programme, IFPRI, ser. 90, 2016. 90.

Belete, A. “Development of Agriculture in Ethiopia since the 1975 Land Reform.” Agricultural Economics, vol. 6, no. 2, 1991, pp. 159–175., doi:10.1016/0169-5150(91)90022-d.

Bunte, Jonas B., et al. “Natural Resource Sector FDI, Government Policy, and Economic Growth: Quasi-Experimental Evidence from Liberia.” World Development, vol. 107, 2018, pp. 151–162., doi:10.1016/j.worlddev.2018.02.034.

Burgis, Tom. “The Great Land Rush: The Billionaire's Farm in Ethiopia.” Investing in Land in Ethiopia, ig.ft.com/sites/land-rush-investment/ethiopia/. 2016.

“Central Statistical Agency Large and Medium Manufacturing and Electricity Industries Survey .” Federal Democratic Republic of Ethiopia Central Statistical Agency, 2012, 2013, 2014, 2015.

Chagas, André Luis Squarize, et al. “A Spatial Propensity Score Matching Evaluation of the Social Impacts of Sugarcane Growing on Municipalities in Brazil.” International Regional Science Review, vol. 35, no. 1, 2011, pp. 48–69., doi:10.1177/0160017611400069.

Cheru, Fantu, and Zinabu Samaro Rekiso. “Ethiopia's 'Economic Diplomacy' and Regional Integration.” The Oxford Handbook of the Ethiopian Economy, by Fantu Cheru et al., Oxford University Press, 2019.

Christiaensen, Luc, et al. “Urbanization and Poverty Reduction - The Role of Rural Diversification and Secondary Towns.” Policy Research Working Papers, 2013, doi:10.1596/1813-9450-6422.

Clapham, Christopher. “Chapter 3: The Political Economy of Ethiopia from the Imperial Period to the Present.” The Oxford Handbook of the Ethiopian Economy, by Fantu Cheru et al., Oxford University Press, 2019, p. 35.

Collier, Paul, and Stefan Dercon. “African Agriculture in 50Years: Smallholders in a Rapidly Changing World?” World Development, vol. 63, 2014, pp. 92–101., doi:10.1016/j.worlddev.2013.10.001.

Cotula, Lorenzo, et al. “Testing Claims about Large Land Deals in Africa: Findings from a Multi-Country Study.” The Journal of Development Studies, vol. 50, no. 7, 2014, pp. 903– 925., doi:10.1080/00220388.2014.901501.

Dell, Melissa, and Benjamin A Olken. “The Development Effects of the Extractive Colonial Economy: The Dutch Cultivation System in Java.” The Review of Economic Studies, vol. 87, no. 1, 2019, pp. 164–203., doi:10.1093/restud/rdz017.

Dercon, Stephen, and Douglas Gollin. “Chapter 26: Agriculture's Changing Role in Ethiopia's Economic Transformation.” The Oxford Handbook of the Ethiopian Economy, by Fantu Cheru et al., Oxford University Press, 2019.

“Development of Agriculture and Agro-Industry in Ethiopia” Stanford Research Institute, 1969.

“Digital Soil Map of the World.” FAO. http://www.fao.org/land-water/land/land- governance/land-resources-planning-toolbox/category/details/en/c/1026564/

Directory of Agriculture. Ethiopian Chamber of Commerce, 1973.

Dorosh, Paul, and James Thurlow. “Agriculture and Small Towns in Africa.” Agricultural Economics, vol. 44, no. 4-5, 2013, pp. 449–459., doi:10.1111/agec.12027.

Dorosh, Paul, et al. “Ethiopia's Spatial and Structural Transformation: Public Policy and Drivers of Change.” IFPRI Ethiopia, EDRI, 2018, www.ifpri.org/publication/ethiopias-spatial-and- structural-transformation-public-policy-and-drivers-change.

“Ethiopia- Elevation Shape File.” UN OCHA. https://data.humdata.org/dataset/ethiopia- elevation-model

“Ethiopian Population Density.” World Pop, 2011-2015. https://energydata.info/dataset/ethiopia- population-density-2015

“Ethiopia: Recent Economic Developments and Prospects for Recovery and Growth.” World Bank, vol. 5929ET, 1987, p. 198.

“Ethiopia Statistical Abstract: Install Generating Capacity and Electric Energy Production by Station.” Central Statistical Office, 1978.

“From Heterogeneity to Homogeneity: Agrarian Class Structure in Ethiopia since the 1950s.” Land and the Challenge of Sustainable Development in Ethiopia: Conference Proceedings, by Dessalegn Rahmato and Taye Assefa, Forum for Social Studies, 2006.

Gabriel, Abebe Haile. Generating Marketed Surplus of Food through State Farms: A Critical Evaluation of the Ethiopian Experience. ser. 72, 1989, Generating Marketed Surplus of Food through State Farms: A Critical Evaluation of the Ethiopian Experience.

Ghose, Ajit Kumar. “Transforming Feudal Agriculture: Agrarian Change in Ethiopia since 1974.” The Journal of Development Studies, vol. 22, no. 1, 1985, p. 132., doi:10.1080/00220388508421972.

Glover, Steven, and Sam Jones. “Can Commercial Farming Promote Rural Dynamism in Sub- Saharan Africa? Evidence from Mozambique.” WIDER Working Paper, 2016, doi:10.35188/unu-wider/2016/164-2.

Gollin, Douglas, et al. “Urbanization with and without Industrialization.” Journal of Economic Growth, vol. 21, no. 1, 2015, pp. 35–70., doi:10.1007/s10887-015-9121-4.

Griffin, Keith, and Roger Hay. “Problems of Agricultural Development in Socialist Ethiopia: An Overview and a Suggested Strategy.” The Journal of Peasant Studies, vol. 13, no. 1, 1985, pp. 37–66., doi:10.1080/03066158508438282.

Hall, Ruth, et al. Africa's Land Rush Rural Livelihoods and Agrarian Change. James Currey, 2015.

Hall, Ruth, et al. “Plantations, Outgrowers and Commercial Farming in Africa: Agricultural Commercialisation and Implications for Agrarian Change.” The Journal of Peasant Studies, vol. 44, no. 3, 2017, pp. 515–537., doi:10.1080/03066150.2016.1263187.

Hausman, J. A. “Specification Tests in Econometrics.” Econometrica, vol. 46, no. 6, 1978, p. 1251., doi:10.2307/1913827.

Heckman, James J., et al. “Matching As An Econometric Evaluation Estimator.” Review of Economic Studies, vol. 65, no. 2, 1998, pp. 261–294., doi:10.1111/1467-937x.00044.

Henderson, B.L. “Master Land Use Plan of Ethiopia.” FAO, 1988. http://www.fao.org/3/ar869e/ar869e.pdf

Henderson, J. Vernon, et al. “The Global Spatial Distribution of Economic Activity: Nature, History, and the Role of Trade.” 2016, doi:10.3386/w22145.

Henderson, J. V., and H. G. Wang. “Aspects of the Rural-Urban Transformation of Countries.” Journal of Economic Geography, vol. 5, no. 1, 2005, pp. 23–42., doi:10.1093/jnlecg/lbh052.

Hirschman, Albert O. “The Strategy of Economic Development.” Yale University Press, 1958, doi:0-300-00559-8.

“HOTOSM Ethiopia Roads (OpenStreetMap Export).” UN OCHA. https://data.humdata.org/dataset/hotosm_eth_roads?force_layout=desktop

“INTEGRATED AGRO-INDUSTRIAL PARKS IN ETHIOPIA.” UNIDO, 2018, www.unido.org/sites/default/files/files/2018-08/Integrated-Agro-Industrial-Parks-in- Ethiopia-Overview-document.pdf.

Kamski, Benedikt, et al. “Omo Investors Won't Scrub Away Kuraz's Sugary Stain.” Ethiopia Insight, 1 Aug. 2019, www.ethiopia-insight.com/2019/08/01/omo-investors-wont-scrub- away-kurazs-sugary-stain/.

Kamski, Benedikt. “The Kuraz Sugar Development Project (KSDP) in Ethiopia: between ‘Sweet Visions’ and Mounting Challenges.” Journal of Eastern African Studies, vol. 10, no. 3, 2016, pp. 568–580., doi:10.1080/17531055.2016.1267602.

Krugman, Paul. “Increasing Returns and Economic Geography.” 1990, doi:10.3386/w3275.

“Large and Medium Scale Commercial Farms Sample Survey.” Federal Democratic Republic of Ethiopia Central Statistical Agency, 2011, 2013, 2014, 2015.

Lavers, Tom, and Festus Boamah. “The Impact of Agricultural Investments on State Capacity: A Comparative Analysis of Ethiopia and Ghana.” Geoforum, vol. 72, 2016, pp. 94–103., doi:10.1016/j.geoforum.2016.02.004.

Lavers, Tom. “‘Land Grab’ as Development Strategy? The Political Economy of Agricultural Investment in Ethiopia.” Journal of Peasant Studies, vol. 39, no. 1, 2012, pp. 105–132., doi:10.1080/03066150.2011.652091.

Main Report: Towards a Strategy for the Development of State Farms in Ethiopia. I, MSFD: Ministry of State Farms Development, 1986, Main Report: Towards a Strategy for the Development of State Farms in Ethiopia.

Mbata, James. “Agribusi Ness i Ndicators: Ethiopia.” World Bank, 68237-ET, Apr. 2012, http://documents.worldbank.org/curated/en/631391468008109813/pdf/682370ESW0P1260 ators0Ethiopia0final.pdf.

Mendes , Diogo Machado, and Lisa Paglietti . Ethiopia Irrigation Market Brief. FAO, 2015, Ethiopia Irrigation Market Brief, www.fao.org/3/a-i5196e.pdf.

“Ministry of State Farms Development Horticulture Development Corporation Annual Report.” Addis Ababa, 1987-1989.

Mirotchie, Mesfin, and Daniel Taylor. “Resource Allocation and Productivity of Cereal State Farms in Ethiopia.” Agricultural Economics, vol. 8, no. 3, 1993, p. 188., doi:10.1016/0169- 5150(93)90013-3.

“National Atlas of Ethiopia.” Ethiopia Mapping Authority, 1981.

National Planning Commission. 2016. Growth and Transformation Plan II (GTP II)—(2015/16– 2019/20). Addis Ababa, Ethiopia: National Planning Commission.

Oqubay, Arkebe. “Industrial Policy and Late Industrialization in Ethiopia.” The Oxford Handbook of the Ethiopian Economy, by Fantu Cheru et al., Oxford University Press, 2019.

Oqubay, Arkebe. Made in Africa: Industrial Policy in Ethiopia. Oxford University Press, 2016.

Pilling, David. “US Ready to Back Ethiopian Reform with $5bn Investment.” Subscribe to Read | Financial Times, Financial Times, 5 Mar. 2020, www.ft.com/content/b0c2963c-5e1a- 11ea-8033-fa40a0d65a98.

Provisional Military Administrative Council (PMAC), 1975. Public ownership of rural lands PMAC Proclam. 31, Addis Ababa

“Regional Atlases of Ethiopia.” Ethiopia Mapping Authority, 1985-1987.

Review and Appraisal of the Financial Performance of State Farms 1981-1985. MSFD: Ministry of State Farms Development, 1987, Review and Appraisal of the Financial Performance of State Farms 1981-1985.

“River Way Shape File.” USGS HydroSheds. https://hydrosheds.cr.usgs.gov/

Roessler, Philip, et al. “On the Origins of Spatial Inequality in Africa.” 2019.

Rosenbaum, Paul R., and Donald B. Rubin. “The Central Role of the Propensity Score in Observational Studies for Causal Effects.” Biometrika, vol. 70, no. I, Apr. 1983, pp. 41– 55., doi:10.1017/cbo9780511810725.016.

“Second Five Year Development Plan”. II ed., Imperial Government of Ethiopia, 1957, pp. 2/7– 4/2,

Sharp, Kay, et al. “Commercialisation of Farming in Ethiopia: Which Pathways.” Ethiopian Journal of Economics, XVI, no. I, Apr. 2007, pp. 1–22.

Shiferaw, Admasu, and Mekonnen Manyazewal. “Economic Policy and Structural Transformation in Ethiopia.” The Oxford Handbook of the Ethiopian Economy, by Fantu Cheru et al., Oxford University Press, 2019.

Shiferaw, Admasu, et al. “Road Infrastructure and Enterprise Dynamics in Ethiopia.” The Journal of Development Studies, vol. 51, no. 11, 2015, pp. 1541–1558., doi:10.1080/00220388.2015.1056785.

Summary of Commissioned Studies: An Assessment of Operation and Performance of Commercial Farms in Ethiopia. UNDP, 2011, Summary of Commissioned Studies: An Assessment of Operation and Performance of Commercial Farms in Ethiopia, www.undp.org/content/dam/ethiopia/docs/Summary%20of%20commissioned%20studies_ Commercial%20farmers%20(2).pdf.

“Summary of Privitization Status from 1987 EC Up to June 2011 EC .” EPA: Ethiopian Privatization Agency, 2018.

Taffesse, Alemayehu Seyoum. “Chapter 27: The Transformation of Smallholder Crop Production in Ethiopia, 1994-2016.” The Oxford Handbook of the Ethiopian Economy, by Fantu Cheru et al., Oxford University Press, 2019, pp. 467–481.

Tefera, Getamesay, and Prof. Xinhai Lu. “Effectiveness Assessment on Overseas Farm Land Investments in Ethiopia.” Applied Science and Innovative Research, vol. 2, no. 4, 2018, p. 193., doi:10.22158/asir.v2n4p193.

Vandercasteelen, Joachim, et al. “Big Cities, Small Towns, and Poor Farmers: Evidence from Ethiopia.” World Development, vol. 106, 2018, pp. 393–406., doi:10.1016/j.worlddev.2018.03.006.

Wiersinga and Jager. (2009) Business Opportunities in the Ethiopian Fruit and Vegetable Sector. Ministry of Agriculture, Nature and Food Quality

Will, Light. “Despite the Construction of Agricultural Processing Industry Parks, the Availability of Raw Material Remains a Concern.” Ethiopia Reporter, 2 June 2019.

VII. Appendix Table 16: 1SLS County Commercial Farm Production (1) (2) VARIABLES Staple Crop Instrument Horticulture Instrument

Has State Farm County 2.967*** 3.760*** (0.232) (0.163)

Capital per Worker 0.0699*** 0.0683*** (0.0263) (0.0258)

Medium Firm (50-100) -0.105 -0.278* (0.149) (0.146)

Large Firm (100+ workers) -0.104 -0.407*** (0.126) (0.124)

Population Density (5-25) -7.736*** -5.605*** (0.266) (0.249)

Population Density (25-100) -8.374*** -5.992*** (0.335) (0.321)

Population Density (100>) -11.37*** -8.356*** (0.372) (0.363)

Road Density 1.405*** 0.544*** (0.105) (0.109)

Year FE Yes Yes Region FE Yes Yes

Constant 2.976*** 2.218*** (0.382) (0.377)

Observations 9,035 9,035 R-squared 0.274 0.302 Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

Table 17: Plot level Commercial Farm Productivity with Interaction terms (1) (2) VARIABLES Staple Horticulture

Staple Crop State Farm County -0.0107 (0.0991)

Horticulture State Farm County 0.212 (0.144)

Horticulture Plot- Low-Lying Region -1.182*** Interaction (0.227)

Staple Plot- Low-Lying Region 0.500* Interaction (0.256)

(0.257) Plot Area (5-50 hectares) -0.0282 -0.0198 (0.0560) (0.0556) Plot Area (50-100 hectares) 0.185 0.176 (0.114) (0.109) Plot Area (100< hectares) 0.213 0.190 (0.150) (0.140) Road Density 0.190 0.222 (0.176) (0.165) Population Density (.2<) -0.232 -0.249 (0.270) (0.276) Population Density (2<) -0.529 -0.560 (0.353) (0.363) Distance to Capital -0.109 -0.0474 (200-400 km) (0.171) (0.149)

Distance to Capital (>400 km) -0.361 -0.0928 (0.258) (0.205) Irrigation 1.422*** 1.575*** (0.280) (0.252) Improved Seeds 0.452*** 0.415*** (0.136) (0.133) Constant 1.729** 1.431** (0.721) (0.695)

Observations 2,790 2,790 R-squared 0.800 0.800 Note: All Continuous Variables are logged and logged values less than 0 are thrown out. Agroprocessing firms include all food, textile, beverage, and leather processing firms. Woreda clustered standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

Table 18: Naive OLS Firm Labor Productivity

(1) (2) (3) (4) (5) VARIABLES All Firms Agroprocessing Food- Food- Food- Processing Processing Processing

Capital per Worker 0.855*** 0.840*** 0.808*** 0.826*** 0.828*** (0.0163) (0.0302) (0.102) (0.0467) (0.0466)

Commercial Farm Prod. -0.000215 -0.00140 -0.00178 (0.00192) (0.00367) (0.00681)

Medium Firm (50-100) 0.0916** 0.113** 0.285*** 0.182*** 0.184*** (0.0377) (0.0513) (0.101) (0.0542) (0.0546)

Large Firm (100+) 0.123*** 0.0478 0.0225 0.121 0.123 (0.0296) (0.0468) (0.0802) (0.0808) (0.0826)

Population Density (5-25) 0.0357 0.0275 0.417*** 0.114 0.0417 (0.0412) (0.0970) (0.154) (0.104) (0.104)

Population Density (25-100) -0.102* -0.0287 0.253 0.0994 0.0234 (0.0538) (0.130) (0.153) (0.156) (0.157)

Population Density (100+) -0.106 -0.0806 -0.0525 -0.143 (0.0702) (0.140) (0.173) (0.176)

Road Density 0.0664*** 0.0635 0.213* 0.0882* 0.107* (0.0227) (0.0491) (0.119) (0.0484) (0.0544)

Staple State Farm County -0.201* (0.105) Horticulture State Farm 0.0208 County (0.0623)

Constant 2.016*** 2.251*** 2.758* 2.438*** 2.409*** (0.198) (0.398) (1.384) (0.618) (0.612)

Observations 8,791 3,098 901 2,104 2,104 R-squared 0.763 0.716 0.723 0.722 0.721 Note: All Continuous Variables are logged and logged values less than 0 are thrown out except commercial farm production. Agroprocessing firms include all food, textile, beverage, and leather processing firms. Woreda clustered standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

100

Table 19: OLS Firm Labor Productivity by Region

(1) (2) (3) (4) VARIABLES All Firms Agroprocessing Agroprocessing excluding Food- Food Processing

Capital per Worker 0.855*** 0.840*** 0.831*** 0.827*** (0.0163) (0.0302) (0.0355) (0.0467)

Medium Firms (50-100 0.0916** 0.113** 0.0764 0.186*** workers) (0.0376) (0.0514) (0.104) (0.0552)

Large Firms (100+ 0.123*** 0.0481 0.0668 0.126 workers) (0.0296) (0.0468) (0.0824) (0.0810)

Population Density (5- 0.0371 0.0357 -0.101 0.0387 25) (0.0380) (0.0976) (0.123) (0.105)

Population Density (25- -0.100** -0.0197 -0.155 0.0194 100) (0.0494) (0.127) (0.199) (0.158)

Population Density -0.104 -0.0675 -0.0224 -0.151 (100+) (0.0647) (0.133) (0.233) (0.180)

Road Density 0.0661*** 0.0616 -0.115 0.113** (0.0226) (0.0467) (0.107) (0.0545)

Gambella -0.652*** -0.746*** -0.709*** (0.0414) (0.0821) (0.127)

Harari -0.302*** -0.238*** -0.118 -0.268*** (0.0419) (0.0651) (0.110) (0.0875)

Addis Ababa -0.150*** -0.210 0.164 -0.252 (0.0526) (0.139) (0.111) (0.186)

Dire Dawa -0.092*** -0.107*** -0.301*** -0.0499 (0.0293) (0.0298) (0.0509) (0.0396)

Afar 0.230 0.231*** 0.108 (0.179) (0.0467) (0.0715)

101

Amhara -0.103** -0.113 -0.206 -0.0698 (0.0441) (0.0705) (0.168) (0.0687)

Oromia -0.131*** -0.119** -0.152 -0.0957 (0.0370) (0.0513) (0.0955) (0.0613)

Benishangul Gumz 0.0749** -1.231*** -1.170*** (0.0348) (0.0439) (0.0489)

SNNP -0.166*** 0.00464 0.311*** -0.0182 (0.0432) (0.108) (0.0782) (0.119)

Year FE Yes Yes Yes Yes

Constant 2.016*** 2.245*** 2.210*** 2.416*** (0.198) (0.397) (0.447) (0.616)

Observations 8,791 3,098 994 2,104 R-squared 0.763 0.716 0.678 0.721 Note: All Continuous Variables are logged and logged values less than 0 are thrown out. Agroprocessing firms include all food, textile, beverage, and leather processing firms. Woreda clustered standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

102

Table 20: Naïve OLS Firm Input per Worker (1) (2) (3) (4) (5) VARIABLES All Firms Agroprocessing Food- Food- Food- Processing Processing Processing

Capital per Worker 0.907*** 0.917*** 0.752*** 0.877*** 0.877*** (0.0318) (0.0422) (0.103) (0.0639) (0.0626)

Commercial Farm Prod. 0.0134*** 0.0252*** 0.0136 (0.00503) (0.00670) (0.0138)

Medium Firm (50-100) -0.400*** -0.135* 0.0187 -0.0674 -0.0969 (0.0751) (0.0786) (0.138) (0.113) (0.108)

Large Firm (100+) -0.446*** -0.271*** 0.0456 0.0708 0.0396 (0.0674) (0.0761) (0.134) (0.121) (0.120)

Population Density (5-25) 0.225** 0.339 0.945*** 0.205 0.272 (0.108) (0.305) (0.335) (0.319) (0.315)

Population Density (25- 0.316** 0.361 0.545 0.224 0.305 100) (0.130) (0.322) (0.361) (0.406) (0.402)

Population Density 0.483*** 0.512 0.0900 0.233 (100+) (0.155) (0.362) (0.445) (0.438)

Road Density -0.0127 -0.00292 0.271 0.0436 -0.0348 (0.0480) (0.0971) (0.209) (0.119) (0.114)

Staple State Farm County 0.0838 (0.181) Horticulture State Farm 0.242* County (0.143)

Constant -0.159 0.0457 2.218 0.736 0.661 (0.378) (0.559) (1.482) (0.851) (0.812)

Observations 7,886 2,919 875 2,064 2,064 R-squared 0.504 0.526 0.480 0.526 0.527 Note: All Continuous Variables are logged and logged values less than 0 are thrown out except commercial farm production. Agroprocessing firms include all food, textile, beverage, and leather processing firms. Woreda clustered standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

103

Table 21: Woreda Food-Processing Production: Staple vs. Horticulture State Farms (1) (2) (3) (4) (5) (6) VARIABLES OLS OLS Propensity Propensity RE RE Propensity Propensity Prod>0 Prod>0

Staple Crop State Farm County 1.629* 1.595 0.329 (0.959) (1.064) (0.871)

Population Density -0.372 -0.328 1.162 1.079 -0.047 0.338 (0.243) (0.233) (0.935) (0.897) (0.665) (0.701)

Road Density 1.625*** 1.563*** 0.499 0.530 1.103 0.621 (0.293) (0.282) (0.866) (0.868) (0.684) (0.761)

Distance to Capital -0.00228 -0.00180 -0.0108* -0.00733 -0.006 -0.00146 (0.00176) (0.00174) (0.00605) (0.00657) (0.0048) (0.00380)

Distance to nearest town with -1.40*** -1.33*** -1.014* -0.871* -0.1687 -0.0489 electricity in 1978 (0.248) (0.245) (0.548) (0.509) (0.263) (0.227)

Horticulture State Farm County 3.941** 4.009** 1.958*** (1.586) (1.802) (0.730)

Constant 6.558*** 6.388*** 8.997*** 8.289*** (1.836) (1.816) (2.926) (2.574)

Observations 2,736 2,736 512 512 112 112 R-squared 0.192 0.202 0.205 0.233

Note: All Continuous Variables except distance to the capital city are logged. Woreda clustered standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1

104